Support linking with libraries of previously-compiled GPU code

[eyalroz]

Some kernels may want to use code that is precompiled, available from other libraries; a concrete example: The more robust printf-family implementation from here. This is currently not possible; so, let us make it possible. The CUDA API wrappers support programmatic linking, so it should be possible to specify additional objects to link with, at least from the command line. For the mean time, let's not get into issues such as compiled object search paths etc.