fix(tts): allow streamed generation to save audio to disk

[mvdirty]

mvdirty note: This work was specified by me, and I am a human. The changes were implemented, documentation was updated, and this PR description was (except for this note) prepared by my clanker, mvagent. I have pretty deep development experience, but not in python, and this is me dipping my toes in just enough to solve a need. The changes seem pretty clean, and we have performed a good range of both manual and automated testing of them, but let me know if anything needs updating to be more idiomatic of either python or the mlx-audio repo style.

---

Context

mlx_audio.tts.generate supports normal file output and also supports low-latency streaming playback with --stream, but the streaming path skipped file writes entirely. That made it hard to use mlx-audio in workflows that need both immediate listening and a saved file for later playback or downstream processing.

This change adds an explicit --save flag for streaming mode so users can opt into both behaviors at once without changing the existing default behavior of --stream.

Description

The TTS CLI save/stream behavior is controlled in mlx_audio/tts/generate.py.

Previously:

• non-streaming generation wrote files to disk
• --stream queued audio to the live player
• the streaming path bypassed the normal output-writing logic

This change introduces --save for use with --stream and preserves the usual mlx-audio file naming and output handling:

• --stream still streams audio during generation and does not write files by default
• --stream --save writes streamed output to disk after generation completes
• --stream --save preserves the usual numbered segment naming such as audio_000.wav
• --stream --save --join_audio writes one combined file such as audio.wav
• output_path and file_prefix are preserved in the streamed save path
• --save without --stream is rejected with a parser error

Changes in the codebase

• mlx_audio/tts/generate.py

  • add a --save CLI flag
  • reject --save unless --stream is also present
  • buffer streamed audio for later writing while keeping live playback unchanged
  • preserve existing numbered output naming for segmented generation
  • preserve existing join_audio, output_path, and file_prefix behavior when saving streamed output
  • factor the final concatenated write path through a small helper

• mlx_audio/tts/tests/test_generate.py

  • add parser validation coverage for --save
  • add regression coverage for streamed save output naming
  • add regression coverage for multi-segment streamed output
  • add regression coverage for output_path / file_prefix handling with streamed save
  • confirm --stream without --save still does not write files

• README.md

  • add CLI examples for --stream, --stream --save, and --join_audio
  • document default numbered output naming and the --join_audio behavior

Changes outside the codebase

None.

Additional information

• The change is intentionally localized to mlx_audio/tts/generate.py; AudioPlayer remains playback-only.
• Host-side smoke testing confirmed:
  • streamed playback begins quickly
  • streamed audio can be saved to disk
  • output_path, file_prefix, and join_audio combinations behave correctly
  • black --check passed on changed Python files
  • isort --check-only passed on changed Python files
  • pytest -s mlx_audio/tts/tests/test_generate.py passes fully
• Local sandbox validation was partially limited because MLX runtime libraries are not available here (libmlx.so), but:
  • git diff --check passed
  • black --check passed on changed Python files
  • isort --check-only passed on changed Python files
  • the new pytest file reaches collection until the missing MLX runtime boundary

Checklist

• Tests added/updated
• Documentation updated