MOGPR performance analysis & improvements

[jdries]

The current MOGPR implementation can be quite heavy, especially when computing/training a new model per timeseries, but also the basic inference.

The main work for this is being done in the GPy library, which also has a successor based on Tensorflow (GPFlow). Hence the most obvious thing to try first, is porting to this tensorflow based library: https://gpflow.github.io/GPflow/develop/notebooks/advanced/multioutput.html

Current GPy model training: https://github.com/Open-EO/FuseTS/blob/main/src/fusets/mogpr.py#L364 And inference: https://github.com/Open-EO/FuseTS/blob/main/src/fusets/mogpr.py#L389

Note that there's also the idea of avoiding to retrain the model per pixel, which is a separate issue!