Bug in SDRClassifier serialization/deserialization

[natoromano]

Updating integration tests with the SDRClassifier showed an error, likely in SDRClassifierRegion. The test opf/opf_checkpoint_test.NonTemporalMultiStep fails when loading a pre-trained model and comparing with one dynamically trained:

AssertionError: 8 != 1 : Mismatch in predicted values: row 278 of a+b has 8 predictions: 
E       ([(0.5269586552579024, 412.0), (0.03142243291403317, 373.26925807999993), (0.02579078528165298, 353.3712486399999), (0.012091269746045623, 379.3483099999999), (0.010465308012194129, 391.236), (0.007283412279594604, 802.0), (0.007121068269599478, 1103.0), (0.007120728734406587, 851.0)]) and row 28 of b has 1 predictions:
E       ([(1.0, 412.0)])

Fix will be part of PR #3168, which updates examples and tests with the SDR classifier instead of the CLA classifier.'

Update: there are actually a few bugs in object conversions when serializing a SDR classifier using the new method. PR #3168 will address those.