Improvements to multivariate output emulator

[jgomezdans]

It would be advantageous to train multi-angle, spectral output emulators with the same basis functions. Basically, the PCA is done in a fairly large number of multi-angle spectra, including a significant amount of ground (LAI=0) spectra. Additionally, we could use the first two basis functions to be the soil, an then add the actual PCA components (so that we have direct access to the soil spectra).

A second thing is that for a multivariate emulator where all individual emulators have the same input training set, the distance metric is the same, so we can re-use this calculation of the K matrix for all emulators

[jgomezdans]

I guess that the distance metric calculation would require quite a lot of recalculations, as the hyperparameters are brought right in...