tests/std: CUDA tests should try device-side code

[sylveon]

The bug introduced in #1336 slipped by testing because the tests only try to build host-side (CPU) code while the error shows up when building device-side code (GPU). We should try building device-side code as well to avoid this happening in the future (so far, I have found that the PR broke PyTorch, Blender, FFMPEG, OpenPose and the Intel Compiler)

[cbezault]

Charles Milette (@sylveon) In order to have caught this we would have had to actually use one of the affected functions in the device code, right?

[sylveon]

I'm not 100% familiar with CUDA, but it seems like it, because otherwise I'd wager that a ton more things would have broken (it was already pretty bad lol). It probably needs a list of functions we intend to "support" from CUDA device-side and then make a test for those (just build-time, not runtime). Just testing everything from <cmath> at least seems like a good start, unless some Nvidia employee wants to weight in with a more complete list.