Adding dilation and dropout

[jmoso13]

This PR adapts the residual units inside the UNet encoder and decoder blocks to follow more closely the encoder/decoder architecture described in both:

• Soundstream
  and
• EnCodec

Specifically it adapts the internal residual units to follow an exponentially increasing dilation schedule by way of a dilation factor base (usually 3):

• dilation=3^0 -> dilation=3^1 -> dilation=3^2

As well as adapting the second convolutional layer in the residual unit to use dilation=1 and kernel_size=1. This follows the residual units described in the papers above.

To optionally remove dilation from the blocks, use dilation factor of 1:

• dilation=1^0 -> dilation=1^1 -> dilation=1^2

Additionally I have added dropout in the residual units for experimentation after the activation. I tried to adhere as closely as possible to the code style you have laid out, let me know if there are stylistic or other tweaks you'd like to see. Hopefully this is a useful contribution! I have already been experimenting in adding dilation to https://github.com/archinetai/audio-diffusion-pytorch using this update and will push a PR in that repo as well for controlling dilation factor and dropout for UNet use in diffusion.

🙏