Benchmarks

Thanks for adding these @anicusan!

I'm thinking of using these as a starting point for the GPUArrays benchmark suite (if that's okay with you obviously). Should I bother with submitting PRs to this codebase with any more complex fixes/changes? I'll definitely open PRs for things like #42, but would you like me to upstream more elaborate changes? For example, I'm thinking of reformatting to make it easier to pull out the benchmarking data and graph and things like that.

Also, these currently only measure kernel launch time for GPU backends since the kernels are not synchronized. I'm not sure what the current best way to add this while making it easy to change backend. Maybe a helper function that adds a synchronization call after running the passed-in function?

Review the existing benchmark suite and how its GPU backends execute and synchronize kernels. Clarify whether benchmark output should be easier to extract and graph, and determine a backend-independent synchronization approach. The work is done when benchmark timings include kernel execution rather than only launch time and the resulting data remains usable across backends.

Written by the indexing model from the issue text.