feat: wire MTP through SimpleEngine for native speculative decoding

[Thump604]

Summary

• Passes the existing --enable-mtp flag through SimpleEngine → MLXLanguageModel → mlx_lm.stream_generate(mtp=True)
• Uses mlx-lm PR #990's native MTP speculative decoding for Qwen3.5 models
• MTP predicts token t+2 using the model's built-in MTP head — no external draft model needed, ~150MB extra memory

Currently --enable-mtp only affects BatchedEngine (scheduler-based MTP). This PR adds the SimpleEngine path, which is the production path for single-user serving.

Changes

File	Change
cli.py	Pass mtp=args.enable_mtp to load_model(), log MTP status for SimpleEngine
server.py	Accept mtp param in load_model(), forward to SimpleEngine
engine/simple.py	Accept mtp in constructor, pass to MLXLanguageModel, log MTP status
models/llm.py	Accept mtp param, pass mtp=True to mlx_lm.stream_generate()

Benchmark (M2 Ultra, 35B-A3B-8bit, streaming API, greedy)

Config	tok/s	Speedup
Baseline (no MTP)	53.0	1.00x
MTP enabled	71.5	1.35x

mlx-lm CLI benchmarks (all models, greedy):

Model	Baseline	MTP	Speedup
27B-8bit (dense)	20.6	27.1	1.32x
35B-A3B-8bit (MoE)	74.4	82.3	1.11x
122B-A10B-5bit (MoE)	43.0	46.7	1.09x

Requirements

• Model must be converted with MTP weights preserved (mlx-lm PR #990)
• MoE models (35B, 122B) also need the MoE sanitize fix for qwen3_5_moe.py
• Pre-converted models: Thump604/Qwen3.5-27B-MLX-8bit, 35B, 122B

Usage

vllm-mlx serve <model-with-mtp-weights> --enable-mtp

Test plan

• Verified MTP flag flows through to mlx_lm.stream_generate(mtp=True)
• Streaming chat completions work with MTP enabled
• Baseline (no MTP) still works unchanged
• --enable-mtp with --continuous-batching uses existing scheduler MTP path (no conflict)

Partially addresses #56

[Thump604]

Closing — fully subsumed by #171 which adds per-request MLLM+MTP routing on top of the basic MTP wiring introduced here.