fix: roll back MTP verify cache on early abort

[pierre427]

Summary

• roll back the uncommitted target-cache draft suffix if an MTP verify round aborts during the single sync / host materialization boundary
• centralize MTP verify-round rollback so the normal reject path trims the MTP draft cache through the same helper
• add a regression test that injects a materialization fault after target verify advances caches but before fresh accept/reject state is read

Why

The target verify forward appends [current_token, draft_1, ..., draft_K] to the main model cache, while the MTP drafter has already appended the K proposed draft positions to its own cache. If cancellation, an injected fault, or a host materialization error fires before the fresh accept state is safely available, the generator should keep the already-committed current token and drop every uncommitted draft position from both caches.

This is intentionally narrow: normal all-accept behavior is unchanged, and the existing reject behavior now goes through the same combined rollback helper.

Validation

• CPU-forced focused tests: tests/test_mtp_spec_decode.py::test_generator_rolls_back_verify_round_on_early_materialization_abort, ::test_generator_emits_first_token_from_backbone_then_draft, ::test_generator_rejection_path_does_not_count_as_accept — 3 passed
• ruff check vllm_mlx/spec_decode/mtp/generator.py tests/test_mtp_spec_decode.py
• ruff format --check vllm_mlx/spec_decode/mtp/generator.py tests/test_mtp_spec_decode.py
• git diff --check -- vllm_mlx/spec_decode/mtp/generator.py tests/test_mtp_spec_decode.py

Full tests/test_mtp_spec_decode.py was attempted CPU-only in this worktree and is environment-blocked on the local system Python 3.9 / missing optional deps (fastapi, mlx_lm.models.qwen3_5), with the same broader toolchain mismatch outside this diff. Default-GPU validation was intentionally not run because a local Qwen3.5-122B quality gate is active on the machine.