Regression-like metrics broken for regression-like tasks

[neworderofjamie]

Metrics are calculated based on the readout values e.g. https://github.com/genn-team/ml_genn/blob/master/ml_genn/ml_genn/compilers/compiled_training_network.py#L310-L316. However, for regression, the readout used to calculate the loss is "var" which, when readout at the end of the batch, only contains the value from the last timestep. This isn't even the right shape to calculate mean-squared error.

The distinction between metrics as things you observe to determine how well the model is performing vs loss functions as things that drive the learning rule is important so I think should be maintained but I fear metrics might also need some way of adding code