Allow JIT compiled deepfiltermodel3 to work in training

[mattpitkin]

When I try turning on JIT compilation of the deepfilternet3 model during training it fails to compile the model. This seems to be due to two things:

• the forward method in SqueezedGRU and SqueezedGRU_S does not explicitly type hint the h argument with Optional[Tensor];
• when setting the coefs.view in the DF class in multiframe.py it need the dimensions to all be explicit rather than trying to work them out from the expansion of *coefs.shape[2:].

This PR fixes both of these things.

The change in multiframe.py does assume that the coefs tensor always has 4 dimension, so if that's not the case then a slight change will be required.

[Rikorose]

I think it can be assumed that coefs is a 4D tensor. Could you fix the formatting, so I can merge?

[mattpitkin]

I've fixed the black formatting, so hopefully all the CI tests should pass this time,

[Rikorose]

Thanks