Refactor: add standard NN torch models

[sgreenbury]

Following discussion with @cisprague, @marjanfamili, @ContiPaolo, @edwardchalstrey1 - standard NN models (such as simple MLP model) would be useful to have readily importable. This could use the PyTorchBackend.

[ContiPaolo]

We may consider adding standard neural networks (e.g., MLP, LSTM for time series) as additional regression models. While NNs often struggle with extrapolation — a key challenge in modeling physical systems — this could emphasize AutoEmulate’s ability to select alterantive, better performing models.

Including NNs would also allow direct comparisons with many existing data-driven surrogate models, which are often based on neural networks.

Uncertainty estimates can be obtained using ensemble techniques.

[sgreenbury]

As discussed could also consider doing UQ with these models through the final outputs for at a given time step parameterizing a distribution (like a gaussian). This is impl-ed in the CNP output.

An illustrative example demonstrating this with 1D time series data set-up.