Add native MTP for Qwen3.6 MLX models

[ffrappo]

Summary

This PR adds native multi-token prediction for Qwen3.6/Qwen3.5-style MLX checkpoints that include in-checkpoint MTP layers. The MTP heads draft candidate tokens, and the target model verifies those candidates before any token is emitted.

What ships:

• Native-MTP loading and dispatch for the selected Qwen3.6 27B and 35B-A3B MLX model cards.
• Exact target verification for drafted tokens, including route-locked verification for the MoE path.
• Probability-ratio speculative sampling for the request's temperature / top_p / top_k / min_p distribution.
• Product/API metadata so clients can see native-MTP capability and final generation stats report when native MTP ran.
• Dashboard and macOS UI surfacing for native-MTP model metadata.

Behavior

Native MTP is enabled by default for supported single-node model cards. It dispatches only when the card declares native_mtp, the local checkpoint has recoverable MTP weights, and the model is placed on one node. Multi-node placements continue to use the normal path in this PR. Operators can disable the feature with EXO_NATIVE_MTP_ENABLED=0 or the macOS setting.

Correctness

Native MTP keeps the target model in charge of emission:

• Greedy decoding preserves target-greedy token IDs. The selected 27B and 35B local execution targets matched target-greedy for K=1/K=2/K=3, fixed and adaptive, with no first divergence in the recorded 64-token runs.
• Sampling uses speculative probability-ratio acceptance against the reconstructed target and draft distributions for temperature / top_p / top_k / min_p.
• Cache repair/commit handling preserves the target KV/GDN state after accepted and rejected drafts.

Current scope:

• Native MTP is enabled for model cards that explicitly declare native_mtp.default_k / native_mtp.max_k.
• Stateful logits_processors such as repetition/presence/frequency penalties are not routed through native MTP yet.
• K>=4 is not enabled by this PR.

Performance

Broad prompt set, max_prompt_tokens=32, max_tokens=64:

Model	Mode	Mean tok/s	vs MTP off	Acceptance
27B native-MTP	MTP off	17.27	1.00x	n/a
27B native-MTP	K=1	29.56	1.71x	85.7%
27B native-MTP	K=2	34.06	1.97x	75.4%
27B native-MTP	K=3	33.79	1.96x	66.4%
35B-A3B native-MTP	MTP off	85.14	1.00x	n/a
35B-A3B native-MTP	K=1	98.59	1.16x	55.8%
35B-A3B native-MTP	K=2	92.27	1.08x	38.3%
35B-A3B native-MTP	K=3	80.53	0.95x	27.4%

Summary: 27B reaches +97.2% / 1.97x at K=2. 35B-A3B reaches +15.8% / 1.16x at K=1 in the broad sweep, and higher K is not automatically better on that path.

Implementation Notes

• The 35B/MoE GDN setup path rebuilds the target prompt cache and primes the MTP cache in one incremental pass while preserving target/MTP cache equivalence.
• Cache setup and repair paths use hidden-state-only target-body calls where logits are not consumed.
• Accepted-prefix counting stays on the MLX side before verification results are needed on CPU.
• K=1 skips draft-token concatenation.
• MTP draft/cache evaluation can overlap with verifier graph construction; EXO_NATIVE_MTP_ASYNC_DRAFT_EVAL=0 disables this path.

Validation

• uv run basedpyright
• uv run ruff check
• nix fmt
• uv run pytest src -q
• Selected 27B and 35B exactness probes for K=1/K=2/K=3, fixed and adaptive