CUDA: Add compiler detection for CUDA < 7.5
This fixes #17920 (closed).
If the CUDA version macros are not defined, the version is reported as 0.0.0. If CUDA version is set to 0.0.0 for NVIDIA, then run nvcc --version
and use that.
(And due to the source being protected by an if defined
, even if the version isn't detected, it still continues to correctly configure instead of providing a strange non-obvious linker failure.)
Edited by Brad King