Possible replacement for #945

[mschauer]

This is how I would approach #945, just as a start. It basically shows how to get the support business working without making the eltype a type parameter of the support. It helps to define support(d, epsilon) to find the numerical relevant support of unbounded distributions.

I illustrate how the type signature of DiscreteNonParametric can be extended to allow uni- and multivariate nonparametric distributions.

The crux is, that Discrete does not need to imply that the support is only Int. All discrete distributions in the original meaning of the word (outcomes can be enumerated, e.g https://en.wikipedia.org/wiki/Probability_mass_function) can also by of ValueSupport type Discrete. I introduced some errors, so this is really just a sketch.

[richardreeve]

As I've said before, I think we just fundamentally disagree about what constitutes a good design pattern for software. My code falls back to an error if you haven't implemented the correct methods (as the code currently does, and I fix in #945 and #951). Yours falls back to the wrong answer (adding a comment in some places that this "is just a fallback" does not help). When (admittedly a long time ago!) I did my PhD in CS, this was considered a bad idea. I'll look at the rest later.