chore: Support supplying the vocoder model locally

[polarathene]

This removes a dependency upon HuggingFace network request at runtime. The model can instead be provided upfront locally, as is supported by the F5-TTS equivalent method called. This makes it opt-in via the ENV MODEL_DIR_VOCODER being configured and having the Vocos model files F5-TTS requires.

I chose an new ENV as it may not be reliable / consistent to place the model at the HF cache path used:

/root/.cache/huggingface/hub/models--charactr--vocos-mel-24khz/snapshots/0feb3fdd929bcd6649e0e7c5a688cf7dd012ef21/pytorch_model.bin

I think this also depends upon a network request regardless, so this permits running the model if the network is unstable too.

---

Example usage:

mkdir --parents ckpts models/dmo-speech-2 models/vocos-mel-24khz
ln --symbolic ../models/dmo-speech-2/model_1500.pt ckpts/model_1500.pt
ln --symbolic ../models/dmo-speech-2/model_85000.pt ckpts/model_85000.pt

# 271M (GRPO-finetuned duration predictor checkpoint) + 4GB (DMOSpeech checkpoint):
wget -P models/dmo-speech-2 https://huggingface.co/yl4579/DMOSpeech2/resolve/main/model_1500.pt
wget -P models/dmo-speech-2 https://huggingface.co/yl4579/DMOSpeech2/resolve/main/model_85000.pt

# 55MB (Vocos - Neural Vocoder):
wget -P models/vocos-mel-24khz https://huggingface.co/charactr/vocos-mel-24khz/resolve/main/config.yaml
wget -P models/vocos-mel-24khz https://huggingface.co/charactr/vocos-mel-24khz/resolve/main/pytorch_model.bin

# Run custom entrypoint (eg: similar to `src/demo.ipynb`):
MODEL_DIR_VOCODER=models/vocos-mel-24khz uv run main.py

---

NOTE: import IPython.display as ipd seems unintentional and was removed (as also proposed by #4).