Add n-gram speculative fallback for native MTP

[youndukn]

Summary

• Add optional draftless n-gram speculation in the native MTP BatchGenerator path.
• Prefer used n-gram continuations, then repeated prompt n-grams, with native MTP as adaptive fallback on n-gram misses.
• Add model settings/profile fields, prompt-token tracking, focused tests, and a concise benchmark note.

Notes

• Disabled by default.
• Current safe target is greedy long-context roleplay or repeated conversation structure with short drafts.

Tests

• python -m py_compile omlx/model_settings.py omlx/model_profiles.py omlx/engine/batched.py omlx/scheduler.py omlx/patches/mlx_lm_mtp/batch_generator.py tests/test_mlx_lm_mtp_patch.py
• PYTHONPATH=/Users/youndukn/projects/oMLX pytest -q tests/test_mlx_lm_mtp_patch.py::TestModelSettingsMtp
• PYTHONPATH=/Users/youndukn/projects/oMLX pytest -q tests/test_mlx_lm_mtp_patch.py

Benchmark summary

40-turn roleplay benchmark, 320 generated tokens, greedy decoding:

• Plain greedy: 48.72 wall tok/s, 62.67 decode tok/s
• N-gram + target fallback: 67.44 wall tok/s, 101.33 decode tok/s
• N-gram + MTP fallback: 68.85 wall tok/s, 103.87 decode tok/s
• N-gram + adaptive MTP fallback: 69.61 wall tok/s, 104.35 decode tok/s