Source builds need workarounds for sdist, CUDA arches, non-x86 CPU dispatch, and setup.py hardcoding

[RemiArata]

Building CTranslate2 from source currently needs multiple downstream workarounds. Core upstream issues:

1. No PyPI sdist (highest impact)
   CTranslate2 publishes wheels but not source distributions, which forces custom source-resolution logic.
   Ask: Publish sdists to PyPI.

2. Deprecated FindCUDA path breaks Blackwell arch handling (related: Feature request: CUDA 13 support #1933)
   Current FindCUDA/select_compute_arch.cmake behavior drops newer multi-digit CUDA arch majors, requiring manual -gencode fixes.
   Ask: Migrate to FindCUDAToolkit + CMAKE_CUDA_ARCHITECTURES.

3. ENABLE_CPU_DISPATCH breaks on unsupported non-x86/non-arm64 arches
   CPU_ISA_DISPATCH is undefined on targets like ppc64le/s390x.
   Ask: Add support or gracefully disable/fallback on unsupported arches.

4. setup.py hardcodes absolute rpaths (related context: Clang unusual switches wrongly hardcoded in resulting setup.py #1704)
   Absolute paths reduce wheel relocatability.
   Ask: Use $ORIGIN-relative rpaths by default.

5. setup.py hardcodes GCC-specific flags (related: Clang unusual switches wrongly hardcoded in resulting setup.py #1704)
   Flags like -fvisibility=hidden and -fno-openmp-implicit-rpath can fail on clang/amdclang.
   Ask: Gate flags by compiler/toolchain detection.

These fixes would significantly improve portability and reduce downstream patching.