James Schloss

Important note: the `chaos_test` introduced in #24 causes a segfault on AMDGPU, but not on CUDA / CPU. This will take some digging to figure out, so I disabled it...

So I am not quite sure where to put this. In #28 I added a test to run all examples. It will run locally, but something about the julia namespace...

As discussed in #1, it would be a good idea to add another dispatch type on `Hutchinson` for this, simply called `:non-random` or something.

Ah, ok. I vaguely remember now. Sorry, that was my bad. The 0 is the neutral element of addition. I should also have a 1 to deal with the neutral...

I fixed the Base.show issue. I honestly don't remember what I was thinking when I did that! I *do* want to ask about what we want to do with these...

This PR breaks the interface to CUDA, so the buildkite tests will fail unless I point to a specific PR (working on that now). How do I do that with...

Ok, can't quite figure out the oneAPI test failures. It is breaking here: https://github.com/leios/GPUArrays.jl/blob/yoyoyo_rebase_time/test/testsuite/statistics.jl#L66 (and at a similar line above). Here is the compare function: ``` function compare(f, AT::Type{

oneAPI passes the tests now. I don't know why. The best I can figure is that we are somehow no longer triggering https://github.com/JuliaGPU/oneAPI.jl/issues/442 (which seems to be hardware dependent)

 Quick note for performance regressions. I ran all the tests on this branch (blue) and plotted them here against master (orange / red). In general, they are the same...

Ok, it seems like this branch works and is ready for review. There is some overall cleanup left, but I'll do that after.