Add serialization test and fix NumUniqueSeparators

[thehomebrewnerd]

Moves serialization test from TestLSA to PrimitiveT so all primitives are tested for ability to serialize.

Fixes serialization issue with NumUniqueSeparators

Closes #113

[rwedge]

I think the calculate feature matrix calls are extending the total time spent for the unit tests by a lot

Do you think that confirming saving and loading the feature would be sufficient for the test (and skip calculating the feature matrixes)?

[thehomebrewnerd]

Do you think that confirming saving and loading the feature would be sufficient for the test (and skip calculating the feature matrixes)?

I think most of the increased time comes from the Elmo serialization test (at least when I test locally). Testing with CFM might be a little overkill, but if we skip CFM with Elmo I think things will run much faster. Let me test it out.

[thehomebrewnerd]

@rwedge Things seem much faster after this change: e27ee38

Should we go with something like this or just drop the CFM stuff completely?

[rwedge]

@thehomebrewnerd Let's keep CFM (barring ELMO)

I'd like to make a call to the serialized feature's primitive get_function before serialization, since for LSA that's when the sklearn pipeline is initialized

[dvreed77]

lgtm