Add support for Apple Foundation Models

Feature request: Add support for Apple Foundation Models (Apple FM) as an OpenAI-compatible provider in Fabric.\n\nGoals:\n- Allow using a local Apple FM server (default endpoint: http://localhost:1976/v1).
- Make API key optional for local servers.
- Handle non-stream SSE (text/event-stream) responses gracefully (concatenate data: fields).
- Provide documentation for running fm serve and any platform requirements.

Notes:
- This targets Apple macOS Golden Gate 27.x (beta 1+) compatibility.
- Recommend guidance that using --host 0.0.0.0 exposes the server to the local network and should be used with caution; include tmux snippet for running the server persistently.

Background: A working prototype exists in PR #2150 that implements SSE parsing and a concatenated-message fallback; creating this issue to track upstream feature acceptance and follow-up work (CI, docs, tests, LFS decision).

Start by reviewing prototype PR #2150 and the existing provider integration entry points. Verify the local http://localhost:1976/v1 setup, optional API key behavior, and handling of non-stream SSE data fields, then check the requested CI, tests, documentation for fm serve, and LFS decision. Done means these follow-ups are resolved with platform and network-exposure guidance included.

Written by the indexing model from the issue text.