Add Kimi K3 standalone B300 perf workload

[khluu]

This PR was authored with assistance from Codex.

Summary

• Adds an opt-in kimi_k3_b300 workload for the standalone 8×B300 runner.
• Reuses the B300 docker#v5.2.0 profile from stacked PR Add GLM-5.2 standalone B300 vLLM benchmark #50 instead of duplicating that infrastructure.
• Uses the release-qualified TP8 Kimi K3 launch shape: DSpark-7, Marlin MXFP4 MoE, FlashInfer MLA, FP8 KV cache, prefix caching, and a 16K context limit.
• Adds an 8K-input / 1K-output, concurrency-64 vllm bench serve cell with 256 prompts.
• Keeps the workload opt-in and consumes VLLM_IMAGE at trigger time.

This does not duplicate PR #57: that PR adds an MI355X aiperf workload, while this change targets the standalone B300 queue and the existing vLLM bench ingestion path.

Trigger

WORKLOADS=kimi_k3_b300
VLLM_COMMIT=7ce42a9bb874a3491819807870e3efebde2845bd
VLLM_IMAGE=public.ecr.aws/q9t5s3a7/vllm-release-repo:7ce42a9bb874a3491819807870e3efebde2845bd-x86_64

Validation

• python3 .buildkite/test_generate_pipeline.py — 11/11 passed
• Workload parser smoke passed
• Dynamic generation emitted queue b300-8, docker#v5.2.0, and the exact release-image override
• bash -n on the runner shell scripts
• Python compile checks for the generator, parser, and generator tests
• git diff --check

The Docker-plugin path itself previously passed on this runner in perf-eval #393. Live Kimi validation is currently blocked by a host issue: #411 shows host nvidia-smi timing out and a minimal GPU container failing with nvidia-container-cli: ... driver rpc error: timed out. After the B300 host is repaired, this PR still needs an end-to-end run against the image above, including confirmation that PR #50's FlashInfer cubin overlay matches the v0.27.1 image.