Support request-level cache for hybrid sliding-window attention

[shihaobai]

Summary

Gemma4 keeps full-attention KV token-granular and stores sliding-window KV in a request-owned pool. Prefix-cache hits restore window checkpoints through main's shared hybrid-attention interfaces.

• Reserve W slots per request and reuse them during single-token decode. Larger prefill chunks borrow temporary slots and release obsolete slots after all layers finish.
• Support KV-shared layers and CPU-pinned window checkpoints for big/small-page GPU prefix-cache hits. CPU-cache pages pack full KV together with window state.
• Bind sliding prefill/decode state to its own token table. Text-only sliding prefill uses the common Triton backend; image tokens retain Gemma's bidirectional mask. Full attention stays on Triton for 512-dimensional heads.
• Co-locate SlidingWindowCacheConfig and SlidingWindowStateCacheManager. PD, MTP, and quantized KV remain unsupported for Gemma4.

Validation

• 95 focused tests passed on H200: config/CPU-page layout, checkpoint copying, CPU-cache transfer, sliding prefill/decode, request-window wraparound/release/restoration, and full/sliding attention numerical comparisons including head_dim=512.
• Repository pre-commit Black and Flake8 passed.
• Full-model inference and CUDA Graph replay were not rerun for this update. Earlier PR revisions passed E4B TP2 and 31B TP4 prefix-cache comparisons; the E4B graph check required communication optimizations to be disabled due to a FlashInfer lazy-init capture issue.