The docstring to method 'gpflow.models.GPR.predict_f_samples' explains the option of
'full_output_cov' of the method. It says 'if [full_output_cov==] True, draw correlated samples over the outputs'.
It's unclear what correlation is meant exactly.

1. It could be correlation between Ys at different Xs (but it's already controlled with parameter 'full_cov', isn't it?).
2. Also, if num_samples ~=None, it could be correlation between Ys in various samples.
3. Something else.