Ccache support for CUDA/HIP builds

[t-a-james]

When compiling for GPU the preprocessor is run twice for each target: once for the host (CPU) and once for the device (GPU) code. Two object files are generated, then later linked together into a "fat binary" which contains both device- and host-side code.

GPU compile times can be orders of magnitude longer than CPU, so it would be great if ccache supported this pattern.

[jrosdahl]

This sounds very similar to #1131 except that the -arch option is not in play?

@t-a-james: Could you provide example command lines and/or a simple CMake project that can be used for reference?