fix(vllm): account for pipeline-parallel size in per_gpu_throughput

[atnair-amd]

Summary

• per_gpu_throughput in to_client_metrics divided total token throughput by tp alone, undercounting GPU count on any pipeline-parallel run (actual GPUs = tp * pp). pp is now a required keyword-only argument.
• Widened the vLLM Ray-backend regression test to assert pp explicitly (subTest over pp=1,2) so a broken pp passthrough can't silently regress again — the prior version only asserted tp/isl.
• Scope is vLLM only. InferenceX ATOM's to_client_metrics call sites are not updated for the new required pp kwarg and are left broken by this change (out of scope here).
• Added GPU metrics polling to the vLLM suite: test_vllm_inference now takes pre/post-model-load VRAM snapshots and polls amd-smi in the background while the client runs; a new test_gpu_metric test reports one HTML row per GPU metric per sweep cell (peak_gpu_memory_mb, model_load_memory_mb, model_load_s, gpu_bandwidth_util_pct, gpu_compute_util_pct). Threshold-coverage validation (vllm_config_loader.py) now gates gpu.* metrics the same way it already gated client.* metrics. amd-smi is invoked via the container-routed orch.exec/orch.exec_on_head — GPU device files (/dev/kfd, /dev/dri) are passed through into the container, so no host-vs-container bypass routing is needed.

Test plan

• python3 -m unittest discover -s cvs/lib/inference/unittests -p "test_*.py" — 188 tests, only known-out-of-scope failures: test_inferencing_config_loader.py (2 pre-existing, unrelated failures confirmed via git stash against commit 5134be1) and test_inferencex_atom_orch_parse.py (3 failures caused by this PR's pp requirement — ATOM is not being updated here, tracked separately)
• python3 -m unittest discover -s cvs/lib/utils/unittests -p "test_*.py" — 126/126 pass (gpu.py polling/aggregation logic)
• python3 -m unittest discover -s cvs/core/orchestrators/unittests -p "test_*.py" — 57/57 pass
• Mutation-tested: reverting the pp fix in vllm_job.py causes the regression test to fail, confirming it actually guards the fix
• Hardware validation on demo cluster: full vLLM suite run against DeepSeek-R1-0528 in a real tp8pp2 config (8-way tensor-parallel, 2-way pipeline-parallel, 2-node Ray cluster) — 43 passed, 1 skipped, rc=0. Confirms per_gpu_throughput and the rest of the metrics pipeline compute and report correctly with pp=2 on live hardware (total throughput 823.47 tok/s, TTFT/TPOT/ITL all populated in the results table).
• pytest --collect-only against cvs/tests/inference/vllm/vllm.py with a real vLLM config confirms test_gpu_metric collects correctly (5 GPU metric rows per cell) alongside the existing 44 tests, with no collection errors.

Known out-of-scope breakage

InferenceX ATOM's to_client_metrics call sites do not pass pp and will fail after this change (ATOM has no pipeline-parallel concept). This is expected and intentionally left unfixed here — scope is vLLM only.