Allow NaN in batch fitting

Make the torch.mean calls NaN compatible such that regularizers can be computed without introducing zero padding

Search the batch-fitting code for torch.mean calls and trace how their results feed regularizer computation. Check how NaN values and zero padding are currently handled, then verify that regularizers can be computed without introducing zero padding while preserving the existing fitting behavior.

Scritto dal modello di indicizzazione a partire dal testo della issue.