fix(qwen35): size the TP sampling scratch from the decodable vocab

[FeathBow]

Description

Fixes #760

#676 bounded token selection to the tokenizer's decodable vocabulary and sized every logits arena from selection_vocab, but it did not touch tp_executor.rs, where the tensor-parallel worker still built its SampleScratch from the checkpoint's padded vocab_size — a line that predates it, from #627. select_batch requires the scratch width to equal the logits width, because a mismatch would otherwise be a silent kernel out-of-bounds rather than a clean error, so on any checkpoint whose tokenizer stops short of the padded width every tensor-parallel prefill and decode failed with select_batch: logits vocab 248077 != scratch vocab 248320. Qwen3.5-4B is such a checkpoint, so TP serving was broken for it.

The four tests that cover this path — the two hf_golden_gate TP2 logits gates, e2e_scheduler TP2, and serving_tp2 — are all #[ignore] because they need two CUDA devices, NCCL, and weights, which is why CI stayed green.

After the change the only remaining uses of vocab_size in this crate are the narrowing comparison itself and the checkpoint tensor-shape validation, both of which want the padded width. Qwen3 and Kimi have no equivalent narrowing and are unaffected.

Type of Change

• Bug fix (non-breaking change which fixes an issue)