fix: support chunked-prefill prompts in the disaggregated serving handoff

[inureyes]

Closes #197

Problem

BatchScheduler::prefill_request_for_handoff and prefill_text_request_for_handoff rejected prompts longer than --prefill-chunk-size: the serving-role entry drove a single full prefill and then extracted, so it bailed on long prompts. This limited disaggregated serving to short prompts, while long prompts are exactly where prefill/decode disaggregation pays off.

Implementation

• prefill_request_for_handoff now mirrors the execute_prefill dispatch: prompts longer than the chunk size go through the standard chunked machinery (start_chunked_prefill followed by continue_chunked_prefill driven to completion) instead of bailing. The final chunk samples the first token via finish_prefill exactly like a single-node chunked prefill (including the terminal-first-chunk handling from Crash ([squeeze] Cannot squeeze axis 1 with size 0) on 3rd turn of multi-turn chat when prompt-prefix cache is enabled — Qwen3-Coder-30B-A3B-Instruct #179), so the full prompt's KV is in the pool before the handoff frame is serialized. Short prompts keep the existing execute_full_prefill path unchanged.
• The chunked-prefill guard in prefill_text_request_for_handoff is removed (it delegates to the now-chunk-capable entry). A defensive guard rejects entry while another chunked prefill is in progress (the serving-role intake is sequential, so this would be a wiring bug).
• The handoff path is text-only, so the VLM-embeddings full-prefill exemption in the normal dispatch cannot apply.

Verification (M1 Ultra, qwen3-0.6b-4bit, real model)

• New serving_handoff_parity_chunked_prefill_matches_single_node_qwen3: a scheduler with prefill_chunk_size = 16 prefills the ~50-token shared prompt in 4 chunks through prefill_request_for_handoff, ships the frame over the coordinator pair, ingests it, and decodes 16 tokens. The decode stream is byte-identical to an UNCHUNKED single-node run through the same scheduler, which also re-proves chunked == full prefill across the handoff boundary.
• All 5 serving-handoff parity tests green (the existing full-prefill handoff, role-loop TCP parity, and the bind/return smoke tests are unaffected).
• Cold clippy (-D warnings, both feature sets), cargo fmt.