fix weights update issue when training separation model

[clement-pages]

BREAKING: remove finetune_wavlm attributes from the PixIT task, and replace it with walm_frozen in Totatonet separation model, to be consistent with what is done for the segmentation model.

[clement-pages]

Merging this PR should fix #1789

[hbredin]

@Lebourdais suggested that we set wavlm.requires_grad to False in __init__ instead of using the with torch.no_grad logic. I think it is indeed a better idea as it allows to get a better estimate of the number of trainable parameters (i.e. those for which requires_grad is set to True.

@clement-pages can you update your PR to make this change and also update the SSeRiouSS model and diarization task accordingly?

[clement-pages]

@clement-pages can you update your PR to make this change and also update the SSeRiouSS model and diarization task accordingly?

OK, I will update it.