fix some FIXMEs in random, and take 2 to #9174

[rfourquet]

In "random/RNGs.jl", some code was unsafe_wraping a pointer of an Array, which was apparently "completely invalid". reinterpret could work in some case as replacement, but it's not flexible enough, e.g. the following doesn't work: reinterpret(UInt128, UInt32[1, 2, 3]). So we use Ptr directly, which is the dSFMT API anyway, with an UnsafeView<:DenseArray wrapper which makes the code nicer. So this is a re-implementation of #9174 in disguise, and allows some nice speed-up for rand! on small arrays as a result of allocating less (unsafe_wrap allocates), for example a = UInt64[1]; rand!(a) has a speed-up of 85% on my machine.

[rfourquet]

@nanosoldier runbenchmarks("random", vs=":master")

[nanosoldier]

Your benchmark job has completed - no performance regressions were detected. A full report can be found here. cc @ararslan

[ViralBShah]

cc @andreasnoack @simonbyrne for review.

[rfourquet]

@vtjnash even without reviewing in detail, would you mind skimming over this to check that this change indeed fixes the previously invalid code? Basically the rand! "kernel" receives a Ptr (wrapped in an UnsafeView), and other rand!(::Array) methods call to the kernel by sending a pointer the array with a length.

[rfourquet]

I would like to merge this today or tomorrow, to make another PR touching this code. I believe this change is correct, and anyway the old one was "invalid" so it's unlikely to get worse.

[StefanKarpinski]

Please go for it then, @rfourquet!

[rfourquet]

Big thanks for your support Stefan with my PRs !