Optimizations:
- xxHash128 batch hashing with struct.pack (kv_cache_utils.py)
- LRU hash cache expanded to 16384 entries for long context
- AITER unified attention + FP8 MFMA paged attention env vars
- Block size 64 optimal for AITER backend
- CUDA Graph capture sizes pruned to [1,2]

Build: bash build_vllm.sh

Co-Authored-By: Claude <noreply@anthropic.com>