Register as buffer Gaussian ratio estimator parameters

[NoemiAM]

Issue: mean and cov of LogRatioEstimator_Gaussian (here) are not networks parameters, hence they do not get saved in checkpoints and they cannot be loaded back.

Proposed solution: register them as buffers

Relevant: there's no PyTorch functionality for loading checkpoints with buffers of unknown size, e.g. where the buffer size is set during model training (see this issue for a relevant discussion).