Residual + Regularisation learning

[balbasty]

Residual field and learning its precision (as in Simpson et al. 2012) seems to work. Also, the lower bound is fixed now: all KL-divergences are >= 0.

TODO:

• smoothness estimates
• template smoothness prior