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/examples/offline_inference/disaggregated-prefill-v1/run.sh

@@ -0,0 +1,11 @@
+rm -rf local_storage/
+
+if [ -f "output.txt" ]; then
+    rm output.txt
+fi
+
+# The directory of current script
+SCRIPT_DIR=$(dirname "$(readlink -f "$0")")
+
+VLLM_ENABLE_V1_MULTIPROCESSING=0 CUDA_VISIBLE_DEVICES=0 python3 "$SCRIPT_DIR/prefill_example.py"
+VLLM_ENABLE_V1_MULTIPROCESSING=0 CUDA_VISIBLE_DEVICES=0 python3 "$SCRIPT_DIR/decode_example.py"