Generating binary cache has a potential race condition in Multiple GPU runs

[sharpneli]

Currently the kernel binary caching does not check if other MPI threads are using a kernel with the same device name. Therefore it should be changed so that only one MPI thread actually writes the binary file.

[ancahamuraru]

See also #86