sccache incorrectly yields hits for code guarded by #ifndef __CUDA_ARCH__

[trxcllnt]

First reported here: NVIDIA/cccl#3103

sccache is incorrectly yielding cache hits on code guarded by #ifndef __CUDA_ARCH__, e.g.:

#ifndef __CUDA_ARCH__
const auto x = "host-only code";
#endif

If the code inside the #ifndef changes, sccache yields a cache hit when it should've been a cache miss, because nvcc -E injects -D__CUDA_ARCH__ (and likely others like -D__CUDA_ARCH_LIST__) into the host preprocessor invocations.

This bug has existed from before I rewrote how sccache handles nvcc compilations.
Luckily, we can fix this because I rewrote how sccache handles nvcc compilations 🙂.