Curiously clang complained about it, but nvcc didn't.

The original algorithm used space as one of the separators between types
in template parameter lists, so as a result "A< short int >" was parsed
into a list of ("A", "short", "int").

This prevented scaling of GMRES on small grids (up to 240x240)!

We must pass the CUDA_HOST_COMPILER variable manually to use the
compiler configured by the user.

Fortunately it's not used anywhere at the moment...

There is still CudaDeviceInfo::getActiveDevice.

Added timer for GPU synchronization.