[ChristianMurphy]

Fixes #50. This adds update_mastery, which updates one student's mastery after one answer without refitting or replaying their history, the incremental update #55 and #57 ask for. Roster now uses it, and stops changing the shared model.

pyBKT/fit/update_mastery.py adds update_mastery, a NumPy version update_mastery_array, and transition_mastery for an unobserved attempt. All three read an answer probability of 0, such as a slip of 0 on a wrong answer, as 1. The forward pass in EM_fit.py and the C++ E-step do the same. That keeps the results equal to predict.

Roster had three problems this fixes:

• It updated a student by swapping the shared model's pi_0 and prior around a predict call, so an update that raised left the model changed.
• On pure Python, creating a Roster with a multigs model, or adding a student to one, raised a broadcast ValueError. This is Error when using a model trained with multigs=True with the Roster class #50.
• set_model passed multigs=True where a class name goes, so it failed for multigs models on the C++ build and on pure Python. pyBKT 1.4 made the same mistake when it created a Roster, which is the KeyError: True in How can I save the fitted model? #52.

Roster now steps through the answers with update_mastery, and set_model refreshes each student without passing any class.

Roster.update_state also checks the classes passed to it, on multigs models and on models fit with multilearn. Each of these raises a ValueError naming the skill and the classes it expects:

• a missing class
• a class the model wasn't fit with
• fewer classes than answers

That last case is the question from #78: should update_state reuse one class for a list of answers? This PR says no, which is how guess and slip classes already behave. Reusing it instead would be a small change. Multiprior and multipair models also have extra learn classes, but they still use their Default class for every answer, as on master.

Both new test files compare against predict_onestep. tests/test_roster.py also checks that the model doesn't change, and each new error. On master, 8 of its 15 tests fail on the C++ build and 10 on pure Python.

[sebastiannberg]

thanks for pushing this, api for update mastery looks good. Only thing I would raise is that a -1 in update_mastery(..., -1, ...) is treated as correct but probably should not be?