refactor: generalize hybrid attention checkpoint management

[shihaobai]

Summary

Extract the existing linear/full checkpoint flow into reusable hybrid interfaces, without including the Gemma/sliding-window implementation.

• Add HybridAttentionReqManager for request-state initialization, big/small-page snapshots, restoration and model-specific MTP state updates. Conv/SSM access stays inside the linear implementation.
• Move checkpoint storage to common/state_cache_manager/: shared slot allocator and layer storage, with the linear config and buffers together in linear_att.py.
• Generalize request fields, paged radix-cache types, multi-level cache routing and PD state-page dispatch across the repository.
• Preserve checkpoint creation timing, size-first CPU layouts, page matching/eviction boundaries, CPU/disk packing and existing --linear_att_* options. PD I/O structures, page_kind literals and model-specific PD packing/restoration remain identical to main.

No changes to attention computation, Gemma/sliding runtime allocation or the small-page policy in this PR.

Validation

Revalidated the review cleanup at 5f1c96a3 on 2026-09-10 with the same cases and main baseline.

Executed in bsh_dsv4 with Qwen3.5-27B, BF16, TP=1 on H200 GPUs, against pristine main 1eb4810c.

• Normal deployment and 1P1D with NCCL transport: 19 requests per mode, 38 requests / 608 generated token IDs, all exactly identical before and after the refactor. Normal and PD outputs also agree for all cases.
• Covered cold inputs, 1024-token big-page hits, 2816-token small-page hits, concurrent requests and GPU eviction followed by CPU restoration. Both normal and PD runs confirmed GPU hit = 0, CPU hit = 2816 on the restoration case.
• Decode CUDA Graph enabled; chunked prefill and CPU cache enabled (CPU cache on P only in PD mode).
• Existing regression tests: 158 passed. The unrelated test_radix_cache.py::test_case10 fails on pristine main as well because its tree has no mem_manager; reproduced separately and excluded from the final selected run.
• pre-commit (Black and Flake8) passed inside the container.

The validation launcher starts P before D to avoid main's existing concurrent NCCL control-port selection race. No unrelated startup fix is included.