feat: connect fresh hybrid KV imports to ATOM PD scheduler #231

Problem

AtomBatcher rejects hybrid Gemma prefill workers even though #225/#227 proved full-prefix RNGD/GPU KV can be imported and continued correctly in a standalone checker.

Evidence

Global layers use assigned KV blocks; sliding layers use lane-owned windows. Current AtomKVBlocks.import_span assumes every layer has global block indexing/full prefix capacity, and staging trusts source dtype instead of actual runtime allocation. Remote admission also must reset the decoder prompt mask.

Scope

Enable fresh full-prefix hybrid PD in the actual AtomBatcher using layer-specific destination/retention and validated physical staging size. Preserve rejection of unimplemented hybrid prefix/host restoration and partial-prefix imports. Add unit ownership/state tests and real Gemma HTTP PD validation, including repeated requests and cancellation/reuse where available.

Done when

The production scheduler accepts a full hybrid KV result, imports correct local/global state, resets only the admitted lane's mask, and matches native output in real PD requests. Unsupported restored/partial paths fail explicitly. Broader prefix/host cache support remains tracked under #201/#207.

Owner

Codex, maintainer review. Related #201, #207, #225, #227.