[EBPF] gpu: replace map with fixed-size array for memory usage tracking

[pjo256]

What does this PR do?

Replaces map[memAllocType]uint64 with [memAllocTypeCount]uint64 fixed-size arrays in the GPU monitoring hot path. This affects two locations:

• kernelSpan.avgMemoryUsage in stream.go — accessed on every CUDA synchronization event
• memTsBuilders in aggregator.go — used when computing per-process GPU memory stats\

If new memory types are added in the future, adding a new constant before memAllocTypeCount in the memAllocType enum is all that's needed.

Motivation

memAllocType is a 4-value enum (kernel binary, global, shared, constant memory). Using a map here adds overhead: hash computation, bucket lookups, and heap allocations on every access. A fixed-size array replaces this with direct indexing.

A benchmarking script shows 6-23x speedup depending on workload (number of kernel launches between syncs), with allocations dropping from 240 B/op to 64 B/op.

Describe how you validated your changes

Existing unit tests in stream_test.go and stats_test.go already index avgMemoryUsage by memAllocType constants (e.g., span.avgMemoryUsage[sharedMemAlloc]), which works for both arrays and maps.

Wrote standalone microbenchmarks reproducing the exact access patterns of getCurrentKernelSpan and getRawStats, comparing map vs array across varying kernel launch counts.

Launches	Map	Array	Speedup (per sync call)
1	260 ns	29 ns	9.0x
10	555 ns	42 ns	13.2x
100	3.4 μs	164 ns	20.5x
500	15.9 μs	700 ns	22.6x
1000	31.5 μs	1.4 μs	23.0x

In a production setting, serving an LLM like Qwen3.5-397B-A17B, we might expect 1500-3000+ kernel launches per forward pass and 10+ forward passes per second - so this can cumulatively save ~20+ms of agent CPU time per second.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[pjo256]

I have read the CLA Document and I hereby sign the CLA

[gjulianm]

Hi @pjo256. From a first look it looks good, I'll run the full test pipeline and will make a full review later.

Thanks a lot!

[pjo256]

@gjulianm Thanks! I've added a reno release note in the latest commit, let me know if anything else needs adjusting.

[gjulianm]

Hi @pjo256, CI is green. Just one minor change, could you remove the changelog note? We don't usually write them for changes like these to avoid having a massive changelog :D

[pjo256]

@gjulianm Done! I saw a failing release-notes check earlier and wasn't sure from the Reno docs 👍