Expose accumulated batched PD producer and transfer durations #249

Problem

Batched Torch PD retains producer prefill_s and transfer_s internally but does not export them. The existing single-stage transfer histogram stays zero on the batched route, so benchmarking cannot distinguish unobserved time from free transfer.

Evidence

#248 Qwen3-0.6B native/PD/DP2 measurements: PD metrics record42remote prefills while airm_kv_transfer_seconds_count remains0. PDBatcher._admit_prefilled accumulates the two durations; BatchedService._pd_metrics exports only counts/tokens.

Scope

Expose cumulative completed-handoff producer compute and transfer durations as distinct batched PD counters, with explicit semantics and missing-stat compatibility. Do not fabricate per-request histograms or infer overlap from summed durations. Add real tiny-model serving tests including repeated scrapes and missing-stat cases.

Done when

Public metrics show nonzero producer compute after actual prefill, transfer totals match the reported handoff durations without increasing on scrape, legacy workers remain supported, and focused/full CPU tests pass. Subsequent performance runs can capture these counters.

Owner

Codex implementation; maintainer review/merge. Related #248 and #47.