[SYCL 2020][USM] Add USM queue shortcuts, memcpy, memset, fill, prefetch, mem_advise stub, test cases

[illuhad]

• Add USM queue shortcuts (e.g. queue::parallel_for)
• Add handler::memcpy
• Add handler::memset
• Add USM handler::fill
• Add handler::prefetch
• Add test cases for all of the above
• Add mem_advise (stub only, see caveats below)

Testing results:

• CPU backend passes all tests
• CUDA backend passes all tests
• HIP backend passes all tests, except memset (see below)

Caveats:

• All the plumbing for prefetch on HIP is there, but the actual call to hipMemPrefetchAsync() is commented out because this function does not yet exist in HIP - likely because HIP doesn't have proper unified memory, so prefetching doesn't make sense.
• It seems hipMemsetAsync() does weird things causing the memset() tests to fail on HIP. This is probably a bug in HIP/ROCm. Since I'm using an older ROCm, it's possible that this is already fixed - needs more investigation.
• mem_advise() is currently only implemented as a stub that throws an exception when actually used.
  • The reason this is a bit more complicated is because the CUDA model doesn't have an async variant for this. This means to get async mem_advise semantics in SYCL we need to launch a host task that then executes cudaMemAdvise(). For this we would first need to implement host tasks, which we don't have yet.
  • Additionally, HIP doesn't have a hipMemAdvise() at all - probably because of a similar reasoning as the missing hipMemPrefetchAsync()
  • On CPU it's not needed because shared allocations on the CPU backend are trivial.