[BugFix] Add patch for local_cache_hit calculation in vllm 0.18.0 met…

[sumingZero]

Purpose

Fix negative local_cache_hit calculation that can crash Prometheus counters.
Related vLLM issue: vllm-project/vllm#36755
When preemption occurs with async scheduling, there's a race condition where:

• schedule(N+1) can preempt a request and reset its state
• update_from_output(N) reads the already-mutated request state
  This can result in num_external_computed_tokens exceeding num_cached_tokens + recomputed, causing local_cache_hit to become negative. Prometheus counters then crash with:
  ValueError: Counters can only be incremented by non-negative amounts.

Modifications

• Patch vllm.v1.metrics.stats.PromptTokenStats.update_from_output to wrap local_cache_hit calculation with max(0, ...)
• Add v0180/vllm/pc/metrics/stats.py with the patched method
• Add v0180/vllm/pc_patch.py to register the patch
• Update apply_patch.py to support vllm 0.18.0 version