Onboarding: offline Dia tests + guide updates for quant predicates, KV cache sizing, and batch contracts

[mm65x]

Summary

Hardens the new-model onboarding path based on three failure modes we hit while porting models recently:

1. Silently broken converted checkpoints when blanket quantization corrupts embeddings tied to output heads or sensitive paths (fixed for fish_qwen3_omni in fix(fish_speech): prevent fast_ path and embeddings from quantizing to fix convert script #905). The adding-a-model guide now documents model_quant_predicate with examples.
2. KV cache overruns when a prompt is prefilled before decoding and the cache was sized only for the generated tokens (fixed for dia in fix(dia): allocate kv cache to fit prefill + generation length #904). The guide now warns to size caches for prefill + generation combined.
3. Batch result contract violations: consumers expect exactly one result per input even when an item produces no audio (fixed for qwen3_tts in fix(qwen3_tts): yield silent audio instead of dropping empty batch items #907). The guide now says to yield silence rather than skip.

Changes

• mlx_audio/tts/tests/test_dia.py (new): first offline tests for Dia — tiny random-weight config, no network/downloads; covers config parsing, generation without ref_audio, and the KVCache capacity contract.
• docs/contributing/adding-a-model.md: new "Prevent Broken Conversions" section (model_quant_predicate), KV cache sizing warning, offline-testing guidance replacing the download-based test example, batch contract note, extended checklist.

Companion PRs carrying their own regression tests: #904, #905.

[lucasnewman]

@mm65x Can you sign your commits and force push so we can clear the merge requirements?