Add support for FUSS/MUSDB separation task

[darius522]

Summary

This PR adds initial support for two new source separation tasks in DASB:

• FUSS (general audio)
• MUSDB18 (music)

These changes include the code required to prepare datasets from their respective raw sources and to compute standard source separation evaluation metrics.

Added Structure

FUSS/

• create_fuss.py: Converts FUSS eval data into a supervised training-compatible format.
• fuss_prepare.py: Builds the DASB-style manifest for FUSS.
• utils.py: Utility functions (file I/O, path mgmt, etc.)
• metrics/bsseval.py: Implements BSSeval metrics (SDR, SIR, SAR).

MUSDB/

• create_musdb.py: Creates chunked training data from original MUSDB train set.
• create_musdb_eval.py: Chunkifies eval/valid splits for supervised testing.
• musdb_prepare.py: Builds DASB-style manifest for MUSDB.
• utils.py: File and audio handling utilities.
• metrics/bsseval.py: Same as FUSS version but task-local for now.

Motivation

These additions bring support for non-speech source separation benchmarks to DASB, expanding its scope beyond Libri2Mix. This lays the foundation for consistent evaluation of discrete audio tokens across music and general audio separation tasks.

Notes

• Evaluation metrics are implemented separately in each task dir but can later be unified if desired.
• No model training or inference logic is included yet; that will follow in a future PR.

Testing

• All data preparation scripts have been run on local machines with subsets of FUSS and MUSDB.
• Output manifests are consistent with Libri2Mix format.
• Evaluation metrics produce valid SDR/SIR/SAR when tested on toy outputs.

Next Steps

• Add train.py and hparams/ for both FUSS and MUSDB.

[pplantinga]

I have updated the main and DASB branch to address the CI failures due to a github change, merging the upstream DASB into this branch should make the CI run properly again

[poonehmousavi]

@darius522 could you please resolve the conflict