Could the Pandas dependency be removed or made optional?

## Description

Pandas is installed as a pip dependency of spleeter. It's adding bloat and complexity, which might not be worth it for for how little it's even used. I don't believe it's referenced at all for a typical "split a song with a pretrained model" workflow.

At a glance, it looks like it could be easily replaced in some instances (e.g. using the [native csv reader](https://docs.python.org/3/library/csv.html) instead of pd.read_csv(), or using dicts instead of dataframes). If there are places where it's really helping things though, perhaps a local import could be wrapped in a try/except, with a friendly error suggesting that the feature requires `pip install pandas==x.x.x`?

Thanks for considering!