[CUDA] runtime wrapper header tweaks

* Pull in host-only implementations of few CUDA-specific math functions.
* #nclude <cmath> early to prevent its inclusion from CUDA headers after
  they've messed with __THROW macro.

llvm-svn: 255933