Capture vllm server log into test_vllm_inference output

[atnair-amd]

Summary

_check_early_failure() only tails the vLLM server log during the
startup/readiness wait; nothing re-checks it afterward, so a
mid-benchmark server crash (e.g. EngineDeadError) never appears in
test_vllm_inference's captured output.

Change

• Added VllmJob.dump_server_log(), mirroring the existing
  dump_client_log() pattern but per-rank, skipping ray-backend
  headless workers (rank > 0), which never run vllm serve
• Wired into test_vllm_inference's failure path, via
  lifecycle.live_server_job so the dump targets whichever job
  actually owns the running server (on the server-reuse path, an
  earlier cell's job -- not the current cell's, which never issued
  build_server_cmd()/start_server())
• Out of scope: no success-path dump -- the server log is per-server,
  not per-cell, so dumping on every passing cell sharing a reused
  server would re-emit the same growing log repeatedly

Tests

• New TestDumpServerLog (3 cases: full content, mp-backend dumps
  every rank, ray-backend skips worker ranks) in
  test_vllm_job_server_reuse.py
• Gate: .test_venv/bin/python -m unittest discover -s cvs/lib/inference/unittests -p "test_vllm_job*.py" (92 passed)
• Full cvs/lib/inference/unittests discovery: 205/207 passed (2
  pre-existing failures in test_inferencing_config_loader.py,
  unrelated, reproduced identically on unmodified dev/dtni)