Fix FET feature selection causing some data to be erased from marginals prediction
With the current implementation, the input data passed to
ClusterCRF.predict_marginals and the output table (theoretically the same table with an extract
p_pred column) would have rows without significant features (the ones determined on fitting) removed. This impacts downstream analysis of precision/recall, since the number of TP,FP,TN,FN is likely affected by this bug.
I also added a test case (which failed before the patch) to make sure the bugfix works as expected and that no regression happens.