Hook into Random API

[KnutAM]

Following Hooking into the Random API I replaced the Base.rand(Type{<:AbstractTensor}) by Base.rand(rng::AbstractRNG, ::SamplerType{<:AbstractTensor}), which has the advantage of supporting

• Using a different generators, e.g. rand(MersenneTwister(1), Tensor{2,3})
• Support rand!(::Array{<:AbstractTensor}).
• Support creating an array of random tensors by e.g. rand(Vec{2}, 2) (However, specific overload is added to get, e.g. Vector{Vec{2,Float64}} in this case instead of Vector{Vec{2}}.)
• Combinations thereof, e.g. rand(MersenneTwister(2), Vec{2}, 2)

AFAIU the same API doesn't work for randn, so this is kept as is.

Julia 1.6: For some reason, the type for apply_all with f(i) = rand(rng, T) (where T<:Number) is not typestable on julia 1.6, but works at least on 1.8 and later. Old implementation is therefore kept to have fast performance for e.g. rand(Tensor{2,3}), but the new functionality in the list above is slow (but typestable output via typeassert) on julia 1.6.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 96.77%. Comparing base (258540f) to head (7b4e1ae).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #216      +/-   ##
==========================================
- Coverage   96.91%   96.77%   -0.14%     
==========================================
  Files          17       17              
  Lines        1490     1521      +31     
==========================================
+ Hits         1444     1472      +28     
- Misses         46       49       +3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[KnutAM]

Ready for review @fredrikekre or @KristofferC