Add Kimi K2.5 PD disaggregation on GB300 Slurm

[khluu]

This PR was authored with assistance from Codex.

Summary

• Add a versioned prefill/decode-disaggregated serving schema to perf-eval workloads.
• Add a Slurm/Pyxis launcher that allocates and supervises vLLM prefill and decode roles plus the vLLM router.
• Add an opt-in Kimi K2.5 NVFP4 workload for three GB300 nodes: one DEP4 prefill node and two independent TEP4 decode nodes.
• Route the workload to the gb300-slurm Buildkite queue and ingest vLLM serving benchmark results with the full serving GPU count.
• Document the workflow, add focused regression coverage, and include a reusable PD-disaggregation model-onboarding agent skill.

Why

perf-eval previously assumed a single local vLLM server. Multi-node PD disaggregation needs one Slurm allocation, role-aware process placement, KV-transfer configuration, router lifecycle management, and benchmark execution inside that allocation. This adds those primitives while keeping the Kimi workload opt-in (nightly: false).

The benchmark loop also now reads workload TSV rows from a dedicated file descriptor. Slurm/Pyxis child processes could otherwise consume stdin and silently skip every benchmark configuration after the first.

Validation

• python3 -m unittest discover -s tests -v — 33 tests passed
• bash -n lib/run.sh lib/server.sh lib/run_lm_eval.sh lib/run_vllm_bench.sh
• python3 -m py_compile .buildkite/generate_pipeline.py lib/ingest_perf.py lib/parse_workload.py lib/slurm_pd_launcher.py
• Buildkite perf-eval #295 passed on the GB300 Slurm cluster, including allocation, all three serving nodes, routing, and the 8K-input smoke benchmark.

Follow-up

Build #295 ran before the dedicated-stdin fix and therefore exercised only the smoke row. A full 8K/1K, concurrency-3072 rerun of the current head remains pending.