Train and save best ARMAConv model

[AgathaSchmidt]

Enhancement description

After we identified the best general model architecture in #1070 and optimized the model with hyperparameter tuning in #1078 we now train the resulting model and save its weight.

This way, the trained weights can be loaded later to make fast predictions on a withholded test dataset.

• model training an saving weights

Checklist

• Attached labels, especially loc:: or model:: labels.
• Linked to project