feat: add mtplx engine support

[outofcoffee]

Proposes the OpenSpec change that adds mtplx as a servable, daemon-supervised, fleet-wakeable engine.

Summary

• mtplx becomes a servable engine: mtplx serve with the model as --model, the served name as --model-id, the context window as --context-window, and --download always passed so the engine fetches a missing model itself
• PARALLEL n renders as --max-active-requests n, MTPLX's admission cap; the scheduling mode stays a preset concern
• The fleet can wake a node for mtplx: the Spinloop-to-deploy-config derivation becomes runner-aware (a node-side runner gate, per-engine preset fallback keys, owned preset keys)
• Readiness via MTPLX's /health; the catalogue gains an mtplx provider; docs and examples/mtplx/

Implementation details

• Cloud deploy is a non-goal: MTPLX is Apple-Silicon-only with no machine image, so remote deploy keeps refusing it and the cloud gate stays llamacpp/vllm.
• Moving the local model-path check to the cloud-only path is a behaviour change beyond mtplx: llamacpp and vllm node wakes with local model files are refused today ("cannot deploy the local model file") and would start working. It is spec'd in the fleet-routing delta.
• No metrics dialect: MTPLX's /metrics is a JSON ring of per-request envelopes, not Prometheus counters, so it gets the omlx treatment — no scrape, no token stats.
• omlx fleet wake is tracked separately (feat: add fleet wake for omlx #159); the runner-aware path makes it a small follow-up.

Refs #148