feat: disaggregated serving-role loops over a real TCP transport (B3a)

[inureyes]

Summary

Epic #116 step #126 B3a. Adds the prefill/decode serving-role loops that drive the disaggregated KV handoff over a transport, and proves them end to end over a real localhost TCP socket.

B2b landed the scheduler handoff entries (prefill_request_for_handoff, ingest_handoff_as_active, decode_handoff_until_idle) and B2c proved them over an in-process MockTransport. B3a wires those entries into role loops on ServingCoordinator and runs them over a real TcpTransport.

Changes

• BatchScheduler::prefill_text_request_for_handoff (src/server/batch/scheduler.rs): builds a queued text sequence from raw request parts (prompt tokens, sampling, max_tokens, response channel, cancel flag) and reuses prefill_request_for_handoff. The empty-prompt and chunked-prefill guards run before allocation, so a rejected request leaks no pool state.
• ServingCoordinator::run_prefill_role / run_decode_role (src/distributed/disaggregated/coordinator.rs): async role loops. Prefill drains requests, prefills each, and ships the extracted frame to the decode peer; decode receives frames, ingests each onto a fresh pool slot, and decodes to completion, streaming tokens on the request channel. The scheduler is borrowed (the coordinator stays model-free for its transport unit tests); the futures are !Send (the scheduler owns the per-process MLX pool), so callers drive them on a current-thread runtime.
• PrefillRoleRequest / DecodeRoleHandoff: the per-request work item and decode coordination metadata. The KV frame carries the cache, prompt history, and generated tokens; the per-request budget, sampling, and output stream stay with the node holding the client connection.

Verification

Real-model two-node parity test (serving_role_loop_parity_matches_single_node_qwen3, #[ignore]): a qwen3-0.6b prefill node and a separate decode node, each its own model and scheduler, connected over a real localhost TcpTransport (decode loop as a concurrent spawn_local task on a current-thread runtime). The prefill node emits the first token and the decode node the continuation, the same split a router merges for a client; the concatenated stream is byte-identical to a single-node run.

reference text: " \n\n###Answer: 4\n\n###Step 1: 2 + "
handoff text:   " \n\n###Answer: 4\n\n###Step 1: 2 + "
  (prefill first token: " \n\n", decode continuation: "###Answer: 4\n\n###Step 1: 2 + ")

Gate: cold clippy -D warnings --all-targets on both feature sets (metal,accelerate and +test-utils), fmt, and the lib unit suites (3081 passed, 0 failed).

Scope / follow-up (B3b)

Additive only; no existing logic changed. The worker flip (a --node-role prefill|decode process drives a role loop) and the two-process RequestRouter/StreamBridge HTTP path are deferred to B3b, where --prefill-peers/--decode-peers supply the peer addresses a single process otherwise lacks. The role-loop futures and channels are shaped for that wiring.

Part of #126.