ZScoreTransform and UnitRangeTransform do not support CuArray

[sdewaele]

ZScoreTransform and UnitRangeTransform do not support CuArrrays. For example, with StatsBase v0.33.2:

using CUDA
using StatsBase

A = 5randn(2,4).+2
t = fit(ZScoreTransform,A;dims=2)

fails with

┌ Warning: Performing scalar operations on GPU arrays: This is very slow, consider disallowing these operations with `allowscalar(false)`
└ @ GPUArrays ~/.julia/packages/GPUArrays/jhRU7/src/host/indexing.jl:43
ERROR: MethodError: no method matching ZScoreTransform(::Int64, ::Int64, ::CuArray{Float32,1}, ::CuArray{Float32,1})
Closest candidates are:
  ZScoreTransform(::Int64, ::Int64, ::Array{T,1}, ::Array{T,1}) where T at /home/student/.julia/packages/StatsBase/EA8Mh/src/transformations.jl:58
Stacktrace:
 [1] fit(::Type{ZScoreTransform}, ::CuArray{Float32,2}; dims::Int64, center::Bool, scale::Bool) at /home/student/.julia/packages/StatsBase/EA8Mh/src/transformations.jl:127
 [2] top-level scope at REPL[4]:1

To fix this, the ZScoreTransform would have to allow AbstractArrays. Also, specifically for CuArrays, to prevent bad performance due to indexing, mean_and_std needs to be updated. See this gist for a possible implementation. I can put together a PR if desired. One question with regards to this change: Is it okay to replace stdm by Statistics.std using the mean option? This will facilitate solving this issue.

See also JuliaGPU/CUDA.jl#426.

[sdewaele]

Fully resolved by PR #622 for ZScoreTransform. UnitRangeTransform has been generalized so it can handle CuArray.

Implement _compute_extrema to make UnitRangeTransform fully functional for CuArray. Note that this implementation is not included in StatsBase because it is not desirable to take on a dependency on CUDA.jl for this.