evaluate predictive uncertatiny over class labels for rotated mnist

[murphyk]

Compute H(p(y|R(x,a)) for increasing angles a, where y={0..9}, when the model is trained on a=0 (standard images). Compare to methods below

https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/DL2/Bayesian_Neural_Networks/Complete_DLII_BNN_2_2.html