fix(vllm): correct heterogeneous prefix-cache hashes

[jeff3071]

Summary

Fix #446

Gemma 4 uses different physical block sizes across KV-cache groups, while request hashes are computed at a shared finer granularity. kvcached previously lost vLLM’s hash_block_size when replacing the native block pool, causing physical blocks to be cached under incorrect prefix hashes.

• Preserves the native vLLM hash_block_size.
• Reuses vLLM’s BlockHashListWithBlockSize for hash conversion.

Testing

• 40 passed in targeted prefix-cache and coordinator tests.
• Tested google/gemma-4-12B-it with vLLM 0.24.0.

[RixinLiu]

LGTM! Thanks for finding the root cause of this bug!