Is your feature request related to a problem? Please describe.

#554 already improved LLPR code a lot, making it work with checkpoints from other models. However it still requires writing a Python script to train a LLPR model, going against the idea of metatrain.

Describe the solution you'd like

We should make LLPR a proper architecture, so that adding uncertainties on a model only requires mtt train llpr-options.yaml

Describe alternatives you've considered

No response

Additional context

No response