Support testing other GPD fitting methods

[sethaxen]

There are a number of methods in the literature for fitting the generalized Pareto distribution. We currently implement both the methods of Zhang & Stephens, 2009 (used in the PSIS paper) and Zhang, 2010, which makes some improvements for k>1. Switching between these two is supported by the improved keyword. This is not well named, as it gives the impression that the Zhang, 2010 method is universally better (for k ∈ (0, 1), the range for which smoothing can help the most, it generally is not). These methods are relatively simple to implement and work well; some others would require heavier dependencies like Optim.jl and may be more costly.

It would be preferable to support easily adding new fitting methods, via something like a gpd_fit_method keyword, which might take a symbol or a singleton object as value. Internally, we should dispatch on this keyword.