The methods did not work with zero array size, so extra checks were
added.

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.