`Wishart` `rand` promotes to `Float64`

[bvdmitri]

The following code should not promote to Float64

julia> Wishart(3f0, [ 1f0;; ]) |> rand
1×1 Matrix{Float64}:
 1.7289491201235807

Works fine for other distributions (I didn't check all of them though):

julia> Normal(0f0, 1f0) |> rand
0.4961709f0

julia> MvNormal([ 0f0 ], [ 1f0 ]) |> rand
1-element Vector{Float32}:
 -0.06129361

logpdf seems to be fine for Wishart:

julia> logpdf(Wishart(3f0, [ 1f0;; ]), [1f0;;])
-1.4189385f0

[devmotion]

The weird ones are actually Normal/MvNormal - generally, samples are not supposed to be related to the type of the parameters, i.e., eltype and partype are different things.

There are a few related issues (e.g., #1071, #960) and even an open (but outdated) PR: #1433 so I'll close this issue as a duplicate.