Some comments on the route to Distributions 1.0

[azev77]

• Make it easy to locate all relevant distributions.
  E.g. suppose I want all continuous univariate distributions w/ support [0, +\infty)...
  MLJ.jl makes this easy:
  using MLJ; X, y = @load_boston;
m = models(): creates a vector of all 132 models.
  m = models(matching(X, y)): vector of 53 models that work w/ the data
  m = models(matching(X, y), x -> x.prediction_type == :deterministic): vec 50 models
  Distributions.jl doesn't currently have the equivalent:
  Distributions.continuous_distributions: arcsine should be Arcsine etc
  The following gets us part of the way there:
  filter(!isabstracttype, subtypes(Distribution))
filter(!isabstracttype, subtypes(UnivariateDistribution))
filter(!isabstracttype, subtypes(MultivariateDistribution))
filter(!isabstracttype, subtypes(MatrixDistribution))
filter(!isabstracttype, subtypes(ContinuousDistribution))
filter(!isabstracttype, subtypes(ContinuousMultivariateDistribution))
  For distributions matching support we discussed: all(insupport.(dist, data))

• it would simplify testing & other things if there was a slightly more structured (almost cookie-cutter) template for adding distributions.
  Some have no default params: Chi(): MethodError: no method matching Chi()
  Sometimes mean= NaN vs mean=Inf
mean(LogitNormal()) gives error (perhaps use numerical?)
  Some dist entropy throws an error instead of NaN or Inf
  Perhaps: The entropy for this distribution has not been coded. Please submit a PR.
  If no closed form entropy is coded/exists, perhaps entropy() should compute it numerically?
  Fit truncated distributions Feature Request: Fit truncated normal #1108
  Some dist don't have quantiles coded: PGeneralizedGaussian, Skellam, VomMises

• A convenient way to loop through all available (non-abstract type) distributions.
  You would find some inconsistencies.
  fieldnames(Normal) gives unicode (:μ, :σ)
fieldnames(Dirichlet) gives (:alpha, :alpha0, :lmnB)

• before 1.0 check out @cscherrer's note.

• this repo is still missing many useful distributions (R Task Views)
  This seems like a great job for a student (maybe JSoC or GSoC or otherwise)

[nickrobinson251]

xref #880