Update GPU CI configuration
CUDA 11.4.1 updates the most recent supported versions of gcc and
clang so we should update those also.
ROCM 4.3 is out, and is more likely than any older version to be the
one we will support in 2022, so we should update them.
Several recent bug fixes in hipSYCL are worth having to avoid needing
to work around them in other changes.
Previously we re-used the hipsycl container for clang-tidy with CUDA,
but as we wish that to remain at clang version 11, we need an alternative
solution. Simplest would be to add CUDA to the linting image, but that
also builds LLVM with TSAN. As that TSAN build didn't work out of the box
on a CUDA base image, I split the TSAN image off from the image that does
linting, and added CUDA to the latter.