Drama speech: accept canonical multi-speaker dialogue and references #1274

Drama's tokenizer and native family prompt builder support speaker IDs 0 through 15, but the speech HTTP request/catalog validator rejects every nonzero speaker. This prevents a request such as <|speaker:0|> Hello. <|speaker:1|> Welcome. from reaching the existing native prompt implementation.

Proposed contribution: admit canonical speaker tags 0 through 15 in requests and reference transcripts, preserve malformed-tag and structural-token refusals, and test exact prompt/reference rows, streaming, and reference-prefix identity. Existing reference passage limits remain in place: sixteen token IDs does not mean sixteen independently cloned voices.

Behavioral reference: Sogni-AI/drama3-mlx dialogue/reference tests. The native change will extend TensorFold's own Apache-2.0 code rather than copy the Fish-derived MLX implementation.

Target: current native main, with separate recipe backports. The older CUDA NVFP4 long-form route additionally needs active-speaker carryover across split segments and topic-shift markers; that adaptation will be tested and scoped explicitly. No kernel arithmetic change or voice-quality claim is proposed. Host correctness and real-model/listening evidence will be distinguished.