Fix CMake scripts so projects using CUDA .cu files build correctly.

[davisking]

This PR makes a few minor changes to the CMake scripts so projects using CUDA .cu files build correctly. For example, CMake allows you to use CUDA files in a python extension module with a CMakeLists.txt such as:

CMAKE_MINIMUM_REQUIRED(VERSION 3.11)
project(pystuff LANGUAGES CXX CUDA)

add_subdirectory(pybind11)

pybind11_add_module(pystuff 
   cpp_stuff.cpp 
   cuda_stuff.cu
   )

That works fine on Linux with the current pybind11 cmake scripts. However, it fails on OS X and windows with Visual Studio. On OS X it fails because of incorrect link visibility settings for the CUDA code. In Visual Studio it fails because the current CMake scripts end up passing /MP and /bigobj to nvcc, but nvcc doesn't understand those options and so it errors out.

This PR fixes those issues. Allowing CMake projects such as the one shown above to compile on all three platforms.

And as an aside, pybind11 is great. I switched dlib to it from boost python and my life is much improved :)

[jagerman]

Seems a reasonable fix to me. One question about:
if(CMAKE_VERSION VERSION_LESS 3.11)
this seems like a very recent version; does it need to be that recent?

[davisking]

I'm not really sure. You need a pretty new version of CMake to compile .cu files in the way I posted earlier, so I don't think you can go much farther back. However, I'm not sure where the exact version threshold is.

[jagerman]

Sure, but I think the actual cmake code itself should work fine when not doing CUDA compilation -- in other words, we can simplify to avoid the if/else branch assuming that the target_compile_options(${target_name} PRIVATE $<$<COMPILE_LANGUAGE:CXX>:${msvc_extra_options}>) generator expression works in older cmake versions.

[davisking]

Right, it doesn't though. I tried a few older versions and cmake errors out since it can't handle the generator expression there.

[jagerman]

Aha, okay, in that case leave it at 3.11. From quick googling it looks like the CUDA first-class language support was added over 3.8--3.10 with the support in earlier cmake versions being rather incomplete/buggy. If someone really wants to be able to use older cmake on Windows for CUDA compilation they can work out a more precise version.

[davisking]

Yep, my thoughts exactly :)

[henryiii]

with the support in earlier cmake versions being rather incomplete/buggy. If someone really wants to be able to use older cmake on Windows for CUDA compilation they can work out a more precise version.

Agreed. And this would be great to have, I've been manually working around this for quite some time.

And, on macOS and Windows, it's easy to keep the CMake version up to date. (Technically, it is easy on Linux too, but since many Linuxes already have old CMake's, a new version is often not added by hand)

[davisking]

Right. So I think this is good to merge, no?

[wjakob]

Looks good, thank you!

[davisking]

Sweet, thanks for merging. And also for making such an awesome library :)