Temporal Memory: Always sort predicted / predicted-active cells

[mrcslws]

When setting the Columns paper code to use the ApicalTiebreakTemporalMemory, experiments that use feedback to L4 will occasionally send unsorted feedforward growth candidates into L2. I see this with multi_column_synapse_sampling. Many SparseMatrix methods expect these to be sorted, and they will fail quietly then cause undefined behavior, unless you have nupic.research.core compiled in Debug mode -- then an assert catches it.

Need to do some due diligence:

• Does this apply to the C++ ApicalTiebreakTemporalMemory?
• Add unit test covering this scenario. The scenario is: some minicolumns have cells predictive with only basal input, while other minicolumns have cells predictive with both basal and apical input. The list of predicted cells and predicted-active cells should remain sorted. (Make sure the test fails without this code change)

[mrcslws]

Due diligence:

• This issue doesn't affect the temporal memory, it just affects consumers of the temporal memory who expect sorted outputs (which it almost always gives). I vaguely remember choosing this behavior on purpose, since the TM doesn't depend on sorted cells, but now it seems more pragmatic to take the slight perf hit and sort them.
• This doesn't affect the C++ ApicalTiebreakTemporalMemory, because the C++ version doesn't operate on batches of cells. There's never a point where the C++ version has a set of "cells in minicolumns where only basal support occurred" and another set of "cells in minicolumns where basal + apical support occurred". Instead, it loops over minicolumns in increasing order and handles them one by one.