Migration towards a python package version.

[nhamilakis]

The goal is to simplify usage of the VTC inference pipeline.

This adds two different modes of execution that use the same code:

Command line :

$ uv run vtc-infer \
    --wavs audios_path \
    --output $output

With (almost) the same arguments as before.

A Python API mode :

from vtc_inference import run_vtc

run_vtc(
    output="results/my_run",
    checkpoint="2.2",
    config="2.2", 
    wavs="data/my_wavs",
)

The model & resource (thresholds & config) is handled fully by huggingface cache, so there is no need to keep those in the repo, their download is automatic and is referenced by the following mapping:

MODEL_TO_REVISION = {
    "2.2": "91167f6",
    "2.1": "91167f6",
    "2.0": "91e67b5",
}

These map to a commit on the model repository.
For backward compatibility, a local path can still be passed instead of a version for config/checkpoint & threshold files.

This pull request is meant to replace #17, which seems to have been a bit forgotten.

[nhamilakis]

I wanted to keep it as backwards compatible as possible, but we could make the checkpoint argument more explicit.
For example, have a --model-version 2.2 and a --checkpoint-path /path/to/best.ckpt that overrides it.