Adds AbstractRNG to various methods which utilise rand.

[simonbyrne]

Partially fixes #10345. Also utilise randexp in randsubseq instead of taking log of rand. I had to move some methods around so that AbstractRNG would be defined when they were parsed.

Remaining methods which call rand without an AbstractRNG argument are

• sprand: this is a bit complicated, as it accepts an optional functional argument for generating the values. now done.
• SharedArray and DArray methods: I don't understand these well enough to know how they interact with the RNG
• mktempdir in file.jl (windows only).
• peakflops in linalg/generic.jl

[ivarne]

DArray seems to be moving (#10334), so a PR to fix the RNG for them should probably go to https://github.com/JuliaParallel/DistributedArrays.jl

mktempdir does uses OS functionality on Unix, so I think relying on the global RNG is fine.

CC: @rfourquet

[simonbyrne]

Yeah, I don't think we need to worry about mktempdir or peakflops.

[simonbyrne]

Okay, I've added some tests, and updated sprand. I think this is ready to go: I think the shared matrix methods can wait until we have a clearer idea of what we're doing with those.

[ViralBShah]

I am ok with this. SharedArray will probably go into a package later on, once we have multi-threading.