adds initial support for orpheus

[mmwillet]

Addresses: #9

Additionally, this adds:

• An abstraction of the general neural audio codec pattern (used by both DAC and SNAC)
• Support for BPE tokenization.
• Fixes a small bug in the #split function
• DRYs up behavior around preparing the output buffer.

The model will be available for testing here

Currently there are some compatibility issues between different GGUF encoders, so preparing the GGUF file locally may require some code changes.

The model has been tested with short, medium and long prompts. It sometimes (though very rarely) has an issue with stopping so using --vad is recommended. Additionally --voice should be set to leah and --temperature to 0.7.

[mmwillet]

Mostly looks fine, but still need to fix python incompatibilities.

[mmwillet]

LGTM