When crosscompiling we pass the sysroot.

We need to try various architecture flags. Clang doesn't automatically
select one that works.  First try the ones that are more likely to work
for modern installations:

* <=sm_50 is deprecated since CUDA 10.2, try sm_52 first for
          future compatibility.
* <=sm_20 is removed since CUDA 9.0, try sm_30.

Otherwise fallback to Clang's current default. Currently that's `sm_20`,
the lowest it supports.

Separable compilation isn't supported yet.

Fixes: #16586