Can not sample from Von Mises

[mbeltagy]

julia> v=VonMises(30,2)
Distributions.VonMises{Float64}(μ=30.0, κ=2.0)

julia> rand(v)
ERROR: MethodError: no method matching quantile(::Distributions.VonMises{Float64}, ::Float64)
Closest candidates are:
  quantile(::Distributions.Distribution{Distributions.Univariate,S<:Distributions.ValueSupport}, ::Real) at /home/elbeltagy/.julia/Distributions/src/univariates.jl:163
  quantile(::Distributions.DiscreteUniform, ::Float64) at /home/elbeltagy/.julia/Distributions/src/univariate/discrete/discreteuniform.jl:114
  quantile(::Distributions.EmpiricalUnivariateDistribution, ::Float64) at /home/elbeltagy/.julia/Distributions/src/empirical.jl:61
  ...
 in rand(::Distributions.VonMises{Float64}) at /home/elbeltagy/.julia/Distributions/src/univariates.jl:64

[matbesancon]

It comes from the fact that the CDF does not have an analytical expression. I wonder if we should not move all such distributions to another package and then offer different routines to find a given quantile.

The other solution is to implement a simple search, we could do that in the mean time

[matbesancon]

Requires #320 to be merged first

[yha]

This is not due to a missing CDF implementation:

julia> v = VonMises(0,1)
VonMises{Float64}(μ=0.0, κ=1.0)

julia> rand(v,1)
1-element Array{Float64,1}:
 1.8643641705208138

julia> rand(v)
ERROR: MethodError: no method matching iterate(::VonMises{Float64})
[...]

The problem is that rand(::UnivariateDistribution,dims...) uses sampler(v), whereas rand(::UnivariateDistribution) bypasses the sampler completely.
This could be fixed by rand(v::VonMises) = rand(sampler(v)), but I'm not sure that's the best fix. Shouldn't the generic rand(d::UnivariateDistribution) refer to the sampler? It could, for example, call _rand(sampler(d)), with the generic _rand implementing the current fallback of using the cdf (currently there _rand! and no _rand).