Higher level recurrent layers

[MartinuzziFrancesco]

Hi! While it is relatively straightforward to implement higher level constructions like LSTM and GRU it's also nice to have a layer provided by the library like PyTorch does, having details like bidirectionality and dropout taken care of.
Are there any plans to include this in Lux as well?

Thanks!

[avik-pal]

Sure I will accept PRs for this. The general layers policy we use is:

1. If a layer is present in Pytorch or Flax, we can have it in Lux (except some cases like Conv2D which is just Conv and such)
2. Boltz.jl is a better place for any other layers.

In this particular case, LSTM should return a AbstractLuxWrapperLayer containing the model (which can be written using Chain I think) and we just do a custom show method for nicer printing.

[MartinuzziFrancesco]

Ok sounds good to me, I'll see if I can get something going!