Add DeepSeek V4 target-only mlx-serve backend

[PhilipJohnBasile]

Summary

• add an exact, immutable 44-shard DeepSeek V4 target-only backend that delegates inference to external mlx-serve
• validate the canonical Hugging Face owner, revision, config, closed shard index, shard/sidecar sizes, and hashes before launch
• add a dedicated macOS app lifecycle for raw mlx-serve health and chat routes instead of treating it as an MTPLX daemon
• refuse to adopt or replace an occupied raw-server port; only stop the process family spawned by this launch
• bypass unsupported MTPLX capabilities, sessions, settings, metrics, cache, thermal, and benchmark endpoints in external mode
• make the UI explicit that this route needs 128 GB, supports target-only AR, and does not provide MTP or DSpark

Verification

• Swift suite: 575/575 passed
• release app build passed
• raw-health lifecycle, occupied-port, non-adoption, spoofed-owner, wrong-revision/config, and missing/extra-shard regressions passed
• focused Python backend tests: 9/9 passed
• live immutable Hub tree matched 44/44 shard names, sizes, and LFS hashes plus all five sidecars (103,855,774,263 bytes)
• Python wheel and sdist built and their archives matched the reviewed source

Boundaries

• this is an external mlx-serve integration, not a native MTPLX model implementation
• MTP and DSpark are intentionally unavailable for this target-only artifact
• model/GPU inference was not rerun in this app-lifecycle patch; the backend retains its memory preflight and does not use --skip-mem-preflight
• Developer ID signing, notarization, Sparkle signing, and the unavailable twine check remain release-environment gates rather than source-merge evidence