[0.1.dev9] Backdoor to add weights and interaction terms for MNL estimation

[smmaurer]

This PR adds a backdoor to the LargeMultinomialLogitStep template so that users can pass their own MergedChoiceTable for model estimation.

The objective is to let people test out the new functionality in ChoiceModels PR #37 (sampling weights, interaction terms) before it's fully integrated into the templates.

Full integration will take a little while because we need to set up UrbanSim data structures and usage guidelines for 2d weights and interaction terms.

Usage

# create a MergedChoiceTable with complicated interaction and sampling
mct = choicemodels.tools.MergedChoiceTable(...)

# set up an MNL model step and give it a model expression (data tables not needed)
m = LargeMultinomialLogitStep()
m.model_expression = 'empl_density + distance_km - 1'

# pass the MergedChoiceTable directly to the fit() method
m.fit(mct)

See this notebook for a demonstration: Interaction-and-sampling-demo-v1.ipynb

Other changes

Separately, this PR also formalizes the utilities for version comparison that we use to make sure saved model steps are recent enough to be loaded into the current version of ModelManager.

These new utilities even have unit tests! 🎉

Versioning

• library version and ModelManager version incremented to 0.1.dev9
• LargeMultinomialLogitStep version incremented to 0.1.dev2