[MLX][Spec] Add Gemma 4 MTP E2E and docs

[wirybeaver]

Stack

PR 6 of 6. Depends on #5.

• Base: prototype/mlx-gemma4-mtp-enable
• Head: prototype/mlx-gemma4-mtp-e2e-docs

Scope

Completes the experimental prototype with:

• minimal JSON-safe /server_info activation and lifecycle counters;
• one pinned real-model exact-token parity gate;
• documented launch flags, limitations, and verification steps.

The test surface deliberately omits the umbrella prototype’s endpoint/horizon matrix, memory soak, reload fingerprints, and unrelated OpenAI response API changes.

Validation

Passed on Apple Silicon:

• targeted pre-commit hooks;
• accumulated offline suite: 54 passed, 65 subtests passed;
• pinned Stage B E2E: 1 passed.

The E2E launches fresh target-only and MTP servers, compares 64 exact cross-window token IDs, validates per-request counter deltas and cleanup, flushes without changing the assistant generation, and verifies an 8-token post-flush prefix.

---

CI States

Latest PR Test (Base): ❌ Run #30053131640
Latest PR Test (Extra): ❌ Run #30053131384