SUMM/Review/Refactor robust

[josef-pkt]

statsmodels.robust never got a make-over. It's still largely the original code.

The basic usecases work quite well, but there are several not commonly used features that have problems or where the code can be improved. We need some PRs with incremental improvements because the big PRs have a tendency to stall. For example, I found bugs while working on enhancements, but then the bugfixes got stuck in the enhancement PRs.

These issues are for fixing and improving the existing code independently of additional enhancements and refactoring needed for extensions.

• RLM with perfect fit: several issues (starting at 55) and one PR
  • robust linear model not converging for very simple case? #2425 (a consequence for nan in linalg?)
• RLM scale options that don't work
• RLM fit starting values ENH: restarting RLM #1354, RLM.fit() does not use start_params #2397
• norms: several bugs in the rho and other methods that are currently not used
  -[x] robust norm Hampel possibly BUGS #1347 Hampel, but there are more bugs
• estimate_location is in norms, what's the appropriate location
• scale: this contains several estimators that are useful as standalone methods, but the code is more like a prototype version, e.g. hardcoded assumptions instead of options, missing convergence criteria or convergence options, recalculating constant numbers inside a loop method.
  • allow maxiter in HuberScale without raising exception #455 what's the status here?
  • robust Huber fails to converge #550
  • robust scale, generalize options for scale-norm #1440
  • scale.Huber is a possible candidate for enhancement makeover because it is the generic joint loc-scale estimator, similar to Huber-Dutter
• cov_type HC, HAC, cluster, ...