Add vLLM v0.18.1 source tree with KV transfer abort fix

third_party/vllm/ now tracked in git for direct patch management.
Based on vLLM v0.18.1 release with one patch applied:

  vllm/v1/core/sched/scheduler.py:
    Replace fatal assert with graceful skip when KV transfer callback
    arrives for an already-aborted request during PD disaggregated serving.

Future vLLM modifications should be made directly in third_party/vllm/
and committed normally. The patches/ directory is kept as documentation
of what changed from upstream.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

third_party/vllm/.buildkite/test_areas/benchmarks.yaml

@@ -0,0 +1,31 @@
+group: Benchmarks
+depends_on: 
+  - image-build
+steps:
+- label: Benchmarks
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/.buildkite"
+  source_file_dependencies:
+  - benchmarks/
+  commands:
+  - bash scripts/run-benchmarks.sh
+
+- label: Benchmarks CLI Test
+  timeout_in_minutes: 20
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+  - pytest -v -s benchmarks/
+
+- label: Attention Benchmarks Smoke Test (B200)
+  device: b200
+  num_gpus: 2
+  optional: true
+  working_dir: "/vllm-workspace/"
+  timeout_in_minutes: 10
+  source_file_dependencies:
+  - benchmarks/attention_benchmarks/
+  - vllm/v1/attention/
+  commands:
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1