Remove all remaining torch calls

[lucasnewman]

The only remaining caller of torch was the Kokoro voice loader, and we can replicate that with a small pickle-based loader to load the single tensor .pt files.

[Blaizzy]

Perfect, indeed it was the only missing dependencies!

However, users will still need torch so transformers can return torch tensors inputs for LLM based TTS models.

So we before we remove it we need to replace return_tensors=mlx or return_tensors=np and test it works.

Ideally with a fallback to torch, such as this:
https://github.com/Blaizzy/mlx-vlm/blob/ef7d6bad980b809a5d4d1e2b778dd736adb865f2/mlx_vlm/utils.py#L851

[Blaizzy]

Seems like all models are fine and return “mlx” specified, expect Sesame:

mlx-audio/mlx_audio/tts/models/sesame/sesame.py

Line 288 in 31c1ff6

text_tokens = self._text_tokenizer.encode(f"[{speaker}]{text}")

[lucasnewman]

Thanks. I also pulled in the load_config() call from mlx-vlm since it's pretty small and avoids some implicit dependencies on mlx-vlm -> transformers -> torch.hub.

[Blaizzy]

LGTM!