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/model_executor.yaml

@@ -0,0 +1,17 @@
+group: Model Executor
+depends_on: 
+  - image-build
+steps:
+- label: Model Executor
+  timeout_in_minutes: 35
+  source_file_dependencies:
+  - vllm/engine/arg_utils.py
+  - vllm/config/model.py
+  - vllm/model_executor
+  - tests/model_executor
+  - tests/entrypoints/openai/test_tensorizer_entrypoint.py
+  commands:
+    - apt-get update && apt-get install -y curl libsodium23
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s model_executor
+    - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py