nvcc: Allow host compiler different to CXX

[kopperp]

CUDA uses the environment variable NVCC_CCBIN to specify the host compiler used by nvcc, e.g., in case of an incompatible system GCC version. Respect this variable in the Makefile.

[fwyzard]

Can't we simple set CXX to a compiler supported by CUDA ?

We try to avoid mixing compilers 🤷🏻‍♂️

[kopperp]

Sure, that also works. You avoid mixing compilers but you need to change CXX when entering/leaving the project.

NVCC_CCBIN can be set globally and is only active for CUDA, with all the upsides and downsides. This MR doesn't really favor any choice, just giving you the option 🤷‍♂️

[kopperp]

Actually, with the current Makefile, setting CXX has no effect since we hardcode to g++. Should we go the CXX route then in a new PR?

[fwyzard]

make CXX=/path/to/another/compiler should work ?

[kopperp]

make CXX=/path/to/another/compiler should work ?

make CXX='g++-15' cuda does indeed work, the CXX environment variable does not.