Native RNG fixes for very large arrays

[maleadt]

The overflow fix isn't great (sacrificing a couple of additional registers), and we probably have a bunch of similar bugs in other native kernels. I contemplated passing a typevar that determines the type of the counter, but that's additional complexity that's probably not worth it (e.g. in randn! the window size depends on the element type, so that would further couple kernel implementation details to the call site).

As noted by @thomasfaingnaert

[maleadt]

base/random: Error During Test at /var/lib/buildkite-agent/builds/gpuci-14/julialang/cuda-dot-jl/test/base/random.jl:201
  Got exception outside of a @test
  Out of GPU memory trying to allocate 4.000 GiB
  Effective GPU memory usage: 99.82% (4.741 GiB/4.750 GiB)
  Memory pool usage: 16.000 KiB (3.188 GiB reserved)

Hmm, I wonder where that 1.6GiB of memory has gone...