Sliceguard

Adapt for better usability on the BengaliAI ASR competition.

* Support MP3s and different sample rates
* Robustify embedding computation
* Adaptive drop reference for large datasets

Special treatment of uni and bivariate detection on nominal (categorical) variables.

Change sliceguard output to return issues in a hierarchical manner. Previously returned a dataframe where each sample could be part of only one issue. Now returns a list of issues where each sample can be part of multiple issues on different hierarchy levels.

This makes the output more aligned with the semantics of min_drop and min_dist and avoids throwing away potential problem samples.

Speed up embedding computation and add outlier filtering.

Show embeddings in spotlight reports by default.

Allow choosing HF hub models for embedding generation and improve use of precomputed embeddings.