Add multi-speaker support to the audio metrics

[quancs]

🚀 Feature

Permutation invariant metrics

Motivation

Multi-speaker tasks like speech separation need to find the best permutation of the speakers in target and the speakers in prediction, and also need to compute metrics in a permutation invariant way.

Pitch

Implement the widely used Permutation Invariant Training (PIT) method

Alternatives

Additional context