Remove bundled FindCUDA.cmake

We bundled FindCUDA so we could suppress warnings in CI testing. In user space, we always use the FindCUDA from the cmake they used, because the tarballs never had these files. But this has lately been confusing while troubleshooting issues.

Instead, we can simply avoid using -Wundef for nvcc CUDA compilation

This trade-off means that CUDA devs can more easily shoot themselves in the foot when accidentally using #ifdef in branches while not using make check-source or native clang CUDA build, but on balance that seems better than bundling FindCUDA.cmake.

Removed outdated workaround for -Wundef with CUDA versions that are no longer supported.