kvcached catches up to vLLM 0.28's Model Runner V2.

Merges on September 24 and 25 (#493, #494, #500) add an attention-only adapter, because V2 bypassed kvcached's hooks by calling a module-level init_kv_cache(). They also add packed four-dimensional K/V storage and hybrid Mamba page binding for V1. V2 still rejects hybrid/Mamba and cross-layer sharing, and ROCm stays on split K/V. 1,504 stars and no tagged release since v0.1.5 in April, so these are main-branch-only.