[src] Beamforming: optimal channel selection

[mpariente]

Based on a posteriori SNR from here

Still need to enable custom channel selection with integer and torch.Tensor

[mpariente]

I'm unsure about the generic get_reference_mic_idx function.
I think it's useful but maybe it's too obscure?

[popcornell]

what if we enforce in every beamforming function that the ref_channel is a vector of shape B, M ?
And get_ref_mics converts integer and vector of B integers to that format ?
In that way one could use a DNN to output the ref_channel like in channel selection. I think it is the most flexible thing

[mpariente]

And could this matrix have soft assignments (as opposed to the one-hot-like we have in this PR)?

[mpariente]

A side note:
I might be wrong but not all beamformers can choose the channel right?
GEV for example doesn't, am I mistaken?