[GPUJIT] Switch from Runtime API calls for managed memory to Driver API calls.

We now load the function pointer for `cuMemAllocManaged` dynamically, so
it should be possible to compile `GPUJIT` on non-CUDA systems again.

It should now be possible to link on non-cuda systems again.

Thanks to Philipp Schaad for noticing this inconsitency.

Differential Revision: https://reviews.llvm.org/D36921

llvm-svn: 311289