[codex] fix(simple): cooperatively cancel specprefill workers

[Thump604]

Summary

• add cooperative cancellation to the SimpleEngine SpecPrefill worker path
• thread a single cancel_check hook through _prefill_draft, score_tokens, and sparse_prefill
• add regression coverage for scoring-time and sparse-prefill cancellation without changing the current main decode shape

Why

_run_blocking_serialized() now keeps the async generation lock held until the worker thread actually finishes, but long-running SpecPrefill phases still need a cooperative exit path after request cancellation. Without that, cancelled SpecPrefill requests can keep burning CPU inside blocking scoring/prefill loops even though the request is already gone.

Validation

• black --fast --check vllm_mlx/engine/simple.py vllm_mlx/specprefill.py tests/test_simple_engine.py
• pytest -q tests/test_simple_engine.py