feat: dedicated disaggregated serving router front (B3b2b)

[inureyes]

Summary

B3b2b of the disaggregated serving productionization (epic #116, capstone #126). Adds a dedicated --node-role router process that completes topology A: client HTTP → router → prefill node (TCP) → decode node (TCP) → router merges (prefill first token + decode continuation) → client SSE. A three-process E2E proves the streamed output is byte-identical to a single-node run.

What this does

• router_front.rs (model-free): a RouterState that loads only a tokenizer + chat template (never model weights), a background demux task routing each ResultFrame to its per-request channel by request_id, RequestRouter-based prefill selection over a registry of the configured peers, and StreamBridge ordering validation. The router tokenizes with the same add_special rule and build_generate_options the model worker uses, so the prompt tokens and sampling match a single-node run. Serves POST /v1/chat/completions (streaming SSE and non-streaming JSON) + /health.
• startup.rs: the "router" node role maps to ServingMode::Router (a router is not a cluster inference role, so no NodeRole variant is added); the serve path branches before the model load to bind the role transport, build the router app, and serve, skipping the model worker and warmup.
• tests/disaggregated_router_e2e.rs: spawns three real mlxcel-server processes (router + prefill + decode) and asserts the streamed SSE content is byte-identical to the single-node greedy reference.

Notes

• The router registers only its prefill/decode peers in the routing registry, never itself, so route_to_prefill cannot select the router's own address. (An earlier draft registered the router as a Hybrid node, which let routing nondeterministically pick the router itself; the real-model SSE E2E caught it.)
• Reuses RequestRouter (node selection) and StreamBridge (phase/ordering validation); the prefill/decode nodes and TCP wire protocol are unchanged from B3a/B3b2a.
• Known follow-ups (documented): the router does not yet apply the chat stream filter (thinking/tool markers pass through, so the E2E uses enable_thinking=false); the decode continuation arrives as one frame and is re-emitted per token (true per-token decode streaming is a later enhancement); only /v1/chat/completions is implemented.

Verification (orchestrator-local, the real metal gate)

• Cold clippy -D warnings --all-targets both feature sets (metal,accelerate and +test-utils): clean
• cargo fmt --check: clean
• 3-process router E2E (tests/disaggregated_router_e2e.rs): router + prefill + decode reproduce "2 + 2 = 4." over SSE, byte-identical to a hybrid single-node run
• Full lib suite: 3086 passed, 0 failed

Part of #126