Forcing users to build X10 from source in order to get CUDA support is not ideal.
Instead we should restructure the x10rt and NativeX10 runtime layers so that all CUDA-related code can be placed into a separate .so file. The lib would only be loaded if the user actually activates CUDA support (for exampling by setting X10RT_ACCELS to have a value).
If the code is structured this way, we should be able to include the CUDA .so in the pre-built X10 release. If the user has CUDA installed and wants to use it, it just works. If they don't, the library is never loaded.