Switch vllm_bench to standalone vllm-bench Rust binary

[khluu]

Summary

• Swap the vllm bench serve Python CLI (run via docker exec inside the vllm/vllm-openai container) for the standalone vllm-bench Rust binary, which runs on the host and talks HTTP to the served port. JSON schema is identical, so ingest_perf.py needed no logic changes.
• The prebuilt Linux musl binary is auto-downloaded to ~/.cache/perf-eval/ on first use, pinned to v0.1.0. Override with VLLM_BENCH_VERSION or point at a local build with VLLM_BENCH_BIN.
• Drop the SPEED-Bench dataset path from lib/run_vllm_bench.sh — it needed HuggingFace download + docker cp + in-container pandas install, and it was already effectively dead code (every workload uses dataset: random). Also remove the now-unused speed_bench_dataset_subset / speed_bench_category schema fields from parse_workload.py, the run.sh TSV reader, and the four AMD workloads (kimi_k2_5_mi300x, kimi_k2_5_mi355x, gpt_oss_120b_mi355x, deepseek_v4_pro_mi355x) that still carried them.
• README updated to point at vllm-bench and drop the now-irrelevant dataset: speed_bench warning.

Net: lib/run_vllm_bench.sh shrinks from 168 → 100 lines and no longer needs docker exec, docker cp, or any in-container pip installs.

This PR was authored with assistance from Claude Code.

Test plan

• bash -n on every helper in lib/
• Parser stub against all 16 workloads — all parse clean, WORKLOAD_VLLM_BENCH_TSV reduced to 7 columns
• Buildkite build on qwen3_5_h200 with BENCH_ONLY=1 to confirm the binary download + host-side invocation work end-to-end on a real GPU agent