CI testing for memory errors

[bernhardmgruber]

The CUB device histogram tests had explicit checking for out-of-bounds accesses by allocating canary zones before and after temporary storage, filled with a known byte pattern before the CUB algorithm invocation and verified afterwards. These checks were removed as part of the port to Catch2 in #1695.

We could generally investigate ideas about checking for memory errors in the CI, e.g. using dedicated tools like ASAN or compute-sanitizer, or generalize the canary zone approach to larger parts of the testing infrastructure (adding it to the CUB launch wrappers).

Originally posted by @gevtushenko in #1695 (comment)

[alliepiper]

Closing as a duplicate of #1618, which will be addressed in #1879. compute-santizer's --padding arg does exactly what we want here.

[bernhardmgruber]

Awesome!