Performance improvements

[radka-j]

Some things to think about:

• what can we parallelise
• how do we exploit locality, data-reuse, caching etc.

Also, we've already made use of some GPyTorch features but is there anything else to consider.

This might become especially important as we start thinking about ensembles.

[radka-j]

We could also look into snapshot ensembles for training large architectures. This involves training a single model rather than M and after some loss has been reached, taking M equally spaced weight snapshots.