[0.2.dev8] Fully enable "backdoor" MergedChoiceTable functionality for Segmented Large MNL

[mxndrwgrdnr]

As the title states, these changes are meant to enable the user to pass in their own MergedChoiceTable object to use for estimating a segmented large-n multinomial logit in choicemodels. The primary use case is for including interaction terms in a model specification (e.g. individual chooser-alternative distances in a location choice model).

The most questionable change I've made is probably in how the LargeMultinomialLogit model handles an mct object passed in to its fit() method. What I've had to do here is to convert the mct to a dataframe, apply the chooser filters to that df, and then re-cast to an mct object per the new functionality in the PR here. This was the simplest solution I could come up with to the problem of segment-based filtering on a user-constructed MergedChoiceTable object. The alternative would have been to do that filtering in the segmented MNL class. Happy to go that route if you prefer.

Let me know what you think.

In any event, we'll want to merge the PR in choicemodels before merging this one.

[mxndrwgrdnr]

FWIW, here is an example of how I'm estimating a segmented large MNL with interaction terms: https://github.com/ual/bayarea_urbansim/blob/py3/notebooks/hlcm_owner_estimate.ipynb

[coveralls]

Coverage decreased (-0.4%) to 91.78% when pulling 2767215 on segment_mct into 1d15d38 on master.

[smmaurer]

@mxndrwgrdnr This all looks good to me. If you increment the version number, i think it's good to merge! I just pushed some commits that fix Travis.