Basic categorical support

[ajratner]

Simple first step:

• Labeling function matrix L now has entries with values in {0,...,K}, where 0 is abstain (and 1 will mean 'nothing' although this is of course immaterial to NS)
• We also pass in the cardinality of L
• Also a redefined Candidate object that has a kwarg for cardinality (+ all the downstream changes needed to Snorkel front-end to support this)

Optionally: Can make the Candidate object fancier by either accepting an integer -- in which case this is the cardinality -- or a list of strings, which will now be the different categorical values

@bach Our thinking is to just support similarity deps for now, sound good?

@thodrek Do you need sparse repr. of L? Discuss with @bryan-he . Also anything else?

TODOs:

• [x] Change compilation code in Snorkel
• [x] Change internal representation of L
• [x] Modify GenerativeModel.marginals for categorical setting
• [x] Write basic tests
• [x] Change Candidate class to have basic support for cardinality > 2
• [x] Store sparse representation of categorical marginals (training and prediction)
• [x] Support for verbose names of categorical values in candidate class
• [x] Add short tutorial on using categorical candidates
• [x] Fix all marginal saving / loading helper functionality
• [x] Extend LogisticRegression end model for categorical
• [x] Extend SparseLogisticRegression end model for categorical
• [x] Fix scoring functions to handle categorical setting

[ajratner]

Merged in dev branch with PR #646; to be merged into master with next release