bug with `deep_ensemble_trajectory` diversify option

[hstojic]

Trajectory sampler at the moment with diversify option samples quantiles from randomly chosen networks and uses that as a tproxy for the trajectory - this is then repeated for the whole batch and idea was to use it to be able to get more trajectories than the networks we have in the ensemble.

This was not correct idea unfortunately, as we end up not drawing from distributions that contain epistemic/estimation uncertainty - this part in deep ensembles comes from combining all the networks. This particularly is problematic with noiseless data where variances of individual network vanish quickly and batch becomes not diverse enough very quickly.

The fix is easy, we should sample trajectories as quantiles from the approximate posterior of the whole ensemble.