vLLM PR #54454: O(1) prefilter for the reasoning-end hot path (structured output)

A per-decode-step host predicate that re-detokenises the WHOLE sequence for every running request, making decoding quadratic in sequence length. is_reasoning_end_streaming runs in EngineCore once per decode step per running request while a structured-output request is still reasoning; the MuseGlimmer implementation answered by decoding the full sequence. Production numbers: MuseGlimmer-30B on RTX PRO 6000 Blackwell, vLLM 0.27.2rc1.dev116+gd4801990a, concurrency 50, JSON schema enforcement on, throughput decaying 260 -> 170 tok/s as sequences grew, GPU drawing ~150 W of its 300 W limit and EngineCore pinned below one CPU core, ~4.9 ms per decode step in that one predicate at 16k reasoning lengths. The fix is a prefilter: check the step's delta token ids against a precomputed set of ids whose text could supply the marker's final character (one pass over the vocabulary, cached per tokenizer), which settles almost every step in O(len(delta)) with no decoding.