[PD] Separate KV transfer backends and validate RNGD-to-ATOM continuation #194

Problem

The experimental native PD path combines a build-pinned RNGD DMA hook with Qwen3-specific KV geometry and a CUDA-only receiver. Reuse across decoder devices needs explicit backend boundaries.

Evidence

PR #193 improves native KV handoff. Stage attribution finds 34–51 ms mean handoff, ~10 ms prepared wait, ~1 ms final GPU import. Existing ATOM paths live on optimus-rm and have separate cache-format/runtime contracts.

Scope

Separate physical-run planning from RNGD ABI hooks; separate KV geometry and upload completion/ownership from CUDA operations; retain the validated native wire contract. Integrate and validate RNGD-to-ATOM continuation using available ATOM runtime, with explicit compatibility checks. Do not claim arbitrary model/vendor support or modify network protocols without evidence.

Done when

Existing GPU correctness passes; generic geometry and backend ownership tests pass; RNGD-to-ATOM hardware continuation is measured or a concrete runtime/model compatibility blocker is recorded. Documentation states backend-specific limits and remaining scheduler work.

Owner

Codex implements and verifies; maintainer reviews and merges. Related #192, #136, #132.