fixes and features for SNPE-A (as used in Gabor RF subproject):
* backup-predict (uncorrected for proposal) if returned precision matrices non-p.s.d.
* numerically stable predict, numerically stable predict from MoG-proposals
* new code for splitting components in last round (based on standardized MoG initialization written in BaseDistribution originally for SNPE-B);
* support for Simulation-based Calibration (SNPE-A can meaningfully track calibration against the proposal prior even in later rounds)
new CNN-related features (as used in Gabor RF subproject):
* improved control over CNN architecture (strides & pooling layers, shapes of input-nomralization parameters)
* extra_inputs: features bypass initial convolutional (or recurrent) layers and directly feed into fully connected layers
merging development forks
fixes and features for SNPE-A (as used in Gabor RF subproject):
new CNN-related features (as used in Gabor RF subproject):