Implement robust fitting methods (e.g. Huber regression)

[abhisrkckl]

TOA outliers (and DM outliers in wideband timing) can be robustly dealt with using methods like Huber regression. This is an alternative to running a full Bayesian analysis for outlier rejection as in https://github.com/nanograv/enterprise_outliers .

Three such methods (Huber function, Bisquare, and Welsch) are available in TEMPO2. (See https://academic.oup.com/mnras/article/468/3/2637/3063899)

We should implement Fitter classes for some or all of these methods.

[aarchiba]

I believe that these methods can operate by reweighting the fitting, so it may make more sense to implement this as modifications of the existing fitters - in particular, the fitting algorithms currently delegate the calculation of appropriate steps to another class, and the primary fitting algorithm (how many iterations, which steps to keep and which to discard) would remain unchanged. The process of calculating the steps to consider taking would change, but that is already customized for different kinds of TOA.