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/attention.yaml

@@ -0,0 +1,25 @@
+group: Attention
+depends_on: 
+  - image-build
+steps:
+- label: V1 attention (H100)
+  timeout_in_minutes: 30
+  device: h100
+  source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
+    - vllm/v1/attention
+    - tests/v1/attention
+  commands:
+    - pytest -v -s v1/attention
+
+- label: V1 attention (B200)
+  timeout_in_minutes: 30
+  device: b200
+  source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
+    - vllm/v1/attention
+    - tests/v1/attention
+  commands:
+    - pytest -v -s v1/attention