integration of addtional features from mnonnenm/delfi fork

[mnonnenm]

• support for input with convoluational structure (such as images), and mixed input (conv. and non-conv.)
• improved handling of calibration kernels under defensive importance sampling (calibrates only on samples from proposal, ignores samples from prior).
• major rewrite of CDELFI/SNPE-A: stability of analytical correction, option to separate 'Alg. 1' and 'Alg. 2' as described in Papamakarios & Murray
• linear recognition networks and class for handling them in an iterative likelihood-free inference algorithm.