[FEAT] MSE_diag_loss as default noise for multivariate model + related tests

Current behavior

The default loss/noise for the multivariate model is the scalar mean absolute error. (MSE)

Proposed behavior

Now, the default loss/noise is the diagnonal mean absolute error (MSE_diag_noise) which correspond to one scalar value per feature. Previous experiments show that this loss leads to better results in practice.

Where to start?

There are very few changes (the default noise has been replace in a json file). This change forced changes on the tests related to the multivariate outputs.

Related to Issue #41 (closed)