Add beamforming modules

[mpariente]

This is inspired by the code in nn-gev from Paderborn and François Grondin's code in speechbrain.
Thanks for open-sourcing it.
Also helped by @popcornell for MWF

Left to code:

• GCCPhat with quadratic interpolation (cf François Grondin's code)
• steering vector
• Blind analytic normalization
• GEV vector phase correction

In parallel

• More docs
• Examples
• Tests

Might be interesting for @boeddeker @FrancoisGrondin @faroit

[popcornell]

GCCPhat with quadratic interpolation (cf François Grondin's code).

We might want to do a Localization module for that.
Also, for tests, we can test against the numpy implementation by Padeborn or ESPNet one.

[mpariente]

Makes sense for the localisation, you're right.

By the way, eternal question, should we make a functional version, and module wrappers?
It might be easier to use for functional addicts

[popcornell]

Can you be more specific ?
Module wrappers you mean wrapping an STFT encoder module in the case of GCC-PHAT ?

[mpariente]

Sorry.

Currently, the beamforming code it's only accessible in nn.Modules.
In PyTorch, all ops are first implemented as functions, and then it also provides some class wrappers.

Should we outsource everything in functions, and only make the function calls in the class, so bothb usage are possible?

[mpariente]

/rebase

[mpariente]

@popcornell could you review please?