[TransferEngine][Store] Fix and complete Biren (SUPA) build and device integration

[Dashener2]

SUPA support was only wired into rdma_context and cuda_alike; every other vendor-conditional omitted USE_SUPA, and gpu_vendor/supa.h mapped CUDA names to non-existent supa* symbols, so it never compiled against a real SUPA SDK.

Transfer Engine:

• Discover GPU topology and classify device memory for SUPA (topology.cpp, memory_location.cpp).
• Enable the TCP GPU staging path for SUPA.
• Build examples and the TCP transport test with SUPA.
• Rewrite gpu_vendor/supa.h against the real SUPA API: runtime symbols use the su prefix, driver symbols use sudrv, driver enums use their su* names, and suMemcpyAsync takes the stream before the kind (wrapped to preserve the CUDA argument order).

Mooncake Store:

• Add AcceleratorVendor::kSupa and register the CUDA-like accelerator device.

Build:

• Include the SUPA SDK as a system include directory so its generic headers (e.g. version.h) do not shadow Mooncake headers.
• Discover the runtime library by name (libsupa/libsupa-runtime) instead of hardcoding it, and link the resolved libraries.

Docs: document -DUSE_SUPA and BIREN_HOME in build.md, quick-start.md and supported-protocols.md.

Tested on 8x Biren166M + 8x RoCE: full TE and Store build and link with -DUSE_SUPA=ON against the SUPA SDK (GPU allocation verified at runtime); the default build without USE_SUPA is unchanged.

Description

Biren (SUPA) support existed only as a skeleton: rdma_context.cpp and
cuda_alike.h knew about USE_SUPA, but every other vendor-conditional omitted
it, and gpu_vendor/supa.h mapped the CUDA API onto supa* symbols that do not
exist in the real SUPA SDK (suGetDeviceCount, suMallocDevice, sudrvMemCreate,
...). As a result the -DUSE_SUPA=ON build could not compile, and even if it had,
GPU topology discovery, device-memory classification, the TCP staging path, and
the Store accelerator device would all have fallen back to CPU paths.

This change completes the Biren integration so -DUSE_SUPA=ON builds and works
end to end:

• Transfer Engine
  • topology.cpp: include USE_SUPA in the GPU topology guard so supa:N
    entries are discovered and NIC/GPU affinity applies.
  • memory_location.cpp: classify Biren device memory via cuPointerGetAttributes
    instead of treating it as host memory.
  • tcp_transport.cpp / tcp_transport_session_impl.h: enable GPU-aware
    staging for SUPA.
  • example/* and tests/tcp_transport_test.cpp: build the GPU paths under SUPA.
  • gpu_vendor/supa.h: rewrite against the actual SUPA SDK. Runtime API uses
    the su prefix, driver API uses sudrv, driver enums use their su* names,
    NVIDIA-only fabric/VMM concepts are dropped, and suMemcpyAsync (which takes
    the stream before the kind) is wrapped to keep the CUDA call order.
• Mooncake Store
  • Add AcceleratorVendor::kSupa, register the CUDA-like accelerator device, and
    enable the cuda-like registry path for SUPA.
• Build
  • Expose ${BIREN_HOME}/supa/include as a system include dir: the SDK ships
    generic headers (notably version.h) that otherwise shadow Mooncake's own
    generated version.h.
  • Resolve the runtime library by name (libsupa for older SDKs,
    libsupa-runtime for current ones) and link the resolved library set.
• Docs: add Biren rows to build.md, quick-start.md and
  supported-protocols.md, documenting -DUSE_SUPA=ON and BIREN_HOME.

No related issue; the work aligns with the "Domestic GPU Support" milestone of
the overall roadmap.

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• Mooncake Conductor (mooncake-conductor)
• Reshard (mooncake-reshard)
• Mooncake EP (mooncake-ep)
• Mooncake PG (mooncake-pg)
• Integration (mooncake-integration)
• P2P Store (mooncake-p2p-store)
• Python Wheel (mooncake-wheel)
• Common (mooncake-common)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Tested inside a Biren SUPA dev container on a host with 8x Biren166M (64 GB each),
Hygon C86-4G CPU and 8x RoCE NICs, using SUPA SDK 1.10.

Test commands:

# 1. SUPA build + link (TE, Store, examples, tests)
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
      -DUSE_SUPA=ON -DBIREN_HOME=/usr/local/birensupa/sdk/latest \
      -DWITH_STORE_RUST=OFF
ninja -C build transfer_engine mooncake_store \
      transfer_engine_bench transfer_engine_validator \
      transfer_engine_bench_with_notify tcp_transport_test rdma_transport_test

# 2. Default build regression (no SUPA)
cmake -S . -B build_default -G Ninja -DCMAKE_BUILD_TYPE=Release -DWITH_STORE_RUST=OFF
ninja -C build_default transfer_engine mooncake_store

# 3. Runtime smoke test: allocate GPU memory on a Biren166M via the TCP path
./build/mooncake-transfer-engine/example/transfer_engine_bench \
  --mode=target --metadata_server=P2PHANDSHAKE --local_server_name=127.0.0.1:12345 \
  --protocol=tcp --use_vram=true --gpu_id=0 \
  --buffer_size=$((256*1024*1024)) --duration=8

Test results:

• Unit tests pass

• Integration tests pass (if applicable)

• Manual testing done (describe below)

• -DUSE_SUPA=ON: 0 compiler errors. libtransfer_engine.a and
  libmooncake_store.a link, and all examples/tests link; the binaries resolve
  libsupa-runtime.so.1 at runtime.

• Runtime smoke test: the target allocates a 256 MiB buffer in Biren VRAM
  ("VRAM is used" / "Allocating memory on GPU 0") and the TCP transport starts
  cleanly, confirming the su* API mapping and the new memory_location /
  topology paths compile and run.

• Default build (USE_SUPA=OFF): 0 compiler errors, no regression to
  existing backends.

• Not covered here: RDMA end-to-end transfer could not be exercised inside the
  container because the zrdma/dinghai10e NIC userspace provider is not
  installed there (ibv_devices is empty), and rdma_transport_test /
  tcp_transport_test were built but not executed. A follow-up run on a bare
  host (or an image with the vendor RDMA provider) is recommended before
  merging.

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh
• I have run pre-commit on the files changed in this PR and all hooks pass
• I have updated the documentation (if applicable)
• I have added tests to prove my changes are effective
• For changes >500 LOC: I have filed an RFC issue

Notes: pre-commit run --files ... passes on all changed files
(trailing-whitespace, end-of-file-fixer, check-merge-conflict,
check-added-large-files, clang-format, codespell and cmake-format); the Python
and Rust hooks are skipped because no Python/Rust files are touched. No new unit
tests were added; the change is compile-gated enablement, and the existing
tcp_transport_test / rdma_transport_test / examples now build under
USE_SUPA. The change is ~264 insertions, under the 500 LOC RFC threshold.

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specify below)

An AI coding assistant (opencode, deepseek-flash) was used to locate the missing
USE_SUPA branches, rewrite gpu_vendor/supa.h against the real SUPA SDK
symbols, adjust the CMake, and run the containerized build/regression
validation. All changed lines were reviewed by the submitter, who can defend the
change end to end.