Make RobustDiff work with variable step size and missing data

[pavelkomarov]

Shouldn't be hard. It's essentially the same as what I've already done in the Kalman filter, taking matrix exponentials to step whatever distances and performing a check to step around any NaNs to make sure they don't get passed to the convex solver and wreck things.

I'm not sure whether the code will be quite as easy to vectorize with different $A_n$ for each step. There is some risk this could slow down cvxpy.