Rework nvrtcc

[davebayer]

I wanted to improve nvrtcc for quite some time, I wanted to add support for linking device libraries, so we can test CUDA device runtime functionality.

First I started improving the current implementation, but then got an idea: is there a possibility that we could just use nvcc but hijack the compilation at the right moment to invoke nvrtc instead of nvcc's device compilation?

Actually, I made it work! When compiling CUDA device code, nvcc uses host compiler to preprocess the source files and passes it to cicc which compiles the CUDA code and generates PTX output.

The idea is that we give nvcc a custom cicc binary that overwrites the generated PTX file with the PTX file compiled with nvrtc.

This approach has many advantages, because we basically need't to rewrite the whole compilation pipeline. And on top of that we can only overwrite the PTX and still use all of the other functionality provided by cicc.

So.. we can just simply call all of the kernels from host code (!!!) because nvcc does all of the necessary linking and other magic for us. That means that we can use the nvrtcc with ordinary .cu files (with proper guards for host code) and we can start testing other configurations than just 1 thread and 1 block with nvrtc.

The implementation is not complete yet, I'm still missing some options, but I am curious what do you think about this!