refactor(gemma4): the serving contracts speak from one place

[FeathBow]

Description

Closes #1017

The newcomer admission existed in three copies (the plain path, the lane path and the mixed gather) and the decode emit contract in two (the staged deliver and the sampled rows). prepare_newcomer now owns validation, prefix-cache resolution, reservation and the Scheduled handshake behind a NewcomerOptions row per caller; deliver_decode_row owns the stop/token/length emit ladder; and the split-KV state builds, uploads and lends its launch views from its own impl. One deliberate alignment: the gather path's whole-account door moves from a hand-written availability check onto the same eviction-aware reservation the plain path uses (with eviction off, as the gather always behaved), so the door answers identically everywhere. Behavior is preserved — every option row reproduces its caller's path.

Test Env

Single GPU (sm_89, 48 GiB, x86_64)

Verification

• cargo fmt --all -- --check and git diff --check pass; the Gemma 4 release server builds; all-target Gemma Clippy passes with -D warnings; release lib tests pass (52).
• The admission-family gates pass through the maintainer runner at the exact head: mixed-step-against-serial, the gathered walk against the serial path, the gathered lifecycle, the coalesce-door burst gate, the full-roster pipeline-under-queue gate, the idle-refill fingerprint gate, and the fp8 mixed structural walk — the last driving the refactored gather through mixed_prefill_decode_step on the e4m3 pool, so the reservation alignment is exercised on both storage families.
• The 26B checkpoint answers four greedy public /v1/completions requests (31-, 35-, 33- and 1421-token prompts, 64 forced tokens) with bytes identical to main's binary, md5 219de1e3c09c.