Give users option to choose metric for tuning and model selection

[radka-j]

Currently we use R^2 when tuning the models and then log R^2 and MSE when evaluating model performance.

The users should have an option to choose which metrics to choose. This should include an option to choose metrics that take predictive uncertainty into account (e.g., pick model with best calibration).

[radka-j]

Example of probability aware metrics are calibration or Continuous Ranked Probability Score (CRPS)

[EdwinB12]

Comment from @sgreenbury

One thing that comes to mind on this is the extent of the metrics we'd like to support: in the first instance we could just implement any alternative from torchmetrics? However, supporting probability measures relating to calibration (or more general callables would probably require a bigger change to the API for handling this)

I will start by simply adding torch metrics support