Add Zero-One Inflated Beta MLP

[sgreenbury]

Closes #868.

This PR adds experimental functionality aiming to demonstrate an implementation of an emulator that can be trained on alternative domains (here that's $y \in [0, 1]$):

• A zero-one inflated beta distribution subclassing torch.distributions.Distribution
• Adds experimental emulator subclassing MLP to return a zero-one inflated beta distribution for predictions.

There is a small non-breaking change to the MLP class not in experimental to enable customisable parameters to be specificed for the last layer. Here that's to specify that 5 params per prediction are needed for the zero-one inflated beta.