[https://nvbugs/6262973][feat] Pool NCCL window tensor outputs

[nv-lschneider]

@coderabbitai summary

Description

Reuse engine-owned NCCL-window storage for eligible row-parallel GEMM outputs,
allowing the following all-reduce to consume the GEMM result directly.

The pool keeps a stable allocation for warmup and CUDA-graph capture, normalizes
its capacity across TP ranks, and falls back to the allocating path when a later
shape exceeds the retained capacity.

Performance improvements were observed for NVFP4 Llama 3.3 70B and FP8
QDQ/cuBLAS Llama 3.3 70B. FP8 rowwise Llama 3.1 8B was also exercised.
The cublas_mm path remains explicit opt-in because it did not show a benefit
for GPT-OSS 20B.

Test Coverage

• NCCL-window tensor-pool unit tests, including fixed-capacity fallback.
• TP-wide capacity-selection unit test.
• Llama 3.3 70B NVFP4: exact pool-off/pool-on outputs.
• Llama 3.3 70B FP8 QDQ/cuBLAS: exact outputs with NCCL_SYMMETRIC.
• Llama 3.1 8B FP8 rowwise: exact pool-off/pool-on outputs.
• GPT-OSS 20B cublas_mm: exercised as an explicit opt-in path.

PR Checklist

• PR description explains the change and motivation.
• Relevant tests and model-path validation were run.
• No API change.
• No new dependencies.
• Documentation is not required for this internal implementation change.