support int32 and int64 operations in all benchmarks

Modern GPUs are doing excellent jobs for int32/fp32 but at least with NVidia and possibly AMD are underperforming on int64 which has significant impact on some applications, whether they be fixed point, time keeping, or workhorse types for algorithms such the numeric types reproblas is built on. I can understand in some cases why fp64 is doing bad but int64 has no excuse when int32 performs so many times faster.

Of course maybe these guesses are incorrect, I've only heard of them as heresay.

This issue calls for each benchmark to be extended to support int32/int64 which may help people understand which GPUs will work for what applications better.

Maybe add an option for dividing up which classes you care about, if you want to preserve the existing subset of functionality (like if you didn't want to run 2x as many tests...)

No files or benchmark entry points are named. Start by inventorying the repository's benchmarks and running the existing suite to understand how types are selected. Done means every benchmark supports int32 and int64, with an optional way to preserve the existing subset if needed.

Written by the indexing model from the issue text.