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/performance-benchmarks/tests/genai-perf-tests.json

@@ -0,0 +1,21 @@
+[
+    {
+        "test_name": "llama8B_tp1_genai_perf",
+        "qps_list": [4,8,16,32],
+        "common_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B-Instruct",
+            "tp": 1,
+            "port": 8000,
+            "num_prompts": 500,
+            "reuse_server": false
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "gpu_memory_utilization": 0.9,
+            "max_num_seqs": 512,
+            "dtype": "bfloat16"
+        },
+        "genai_perf_input_parameters": {
+        }
+    }
+]