[CUDA] Fix wrappers for sm_80 functions

Previous implementation provided wrappers for the internal implementations used
by CUDA headers. However, clang does not include those, so we need to provide
the public functions instead.

Differential Revision: https://reviews.llvm.org/D151243