Add metrics for probabilistic outputs

[sgreenbury]

Implementing the following probabilistic metrics would be useful to enable the performance of an emulator's UQ to be quantified.

As mentioned in #517, the following could be good to start:

• CRPS
• Coverage metrics (e.g. proportion times interval captures GT)

[radka-j]

Relates to #847

[radka-j]

See https://lightning.ai/docs/torchmetrics/stable/regression/crps.html

[sgreenbury]

See https://lightning.ai/docs/torchmetrics/stable/regression/crps.html

This looks great for 1D target and for a tensor of predictive samples. I wonder if we can also generalize to:

• y_true when 2D
• DistributionLike as the y_pred type