CXX flags passed to nvcc using CMake make compilation fail

[MarelliF]

Issue description

Compilation fails when using CMake and pybind11_add_module for project containing CUDA code. The -flto flag is passed to NVCC and causes the following output:
nvcc fatal : Unknown option 'flto'

Reproducible example code

In CMakeLists.txt:

cmake_minimum_required(VERSION 3.14)
project(projectname LANGUAGES CXX CUDA)
find_package(pybind11 REQUIRED)

pybind11_add_module(modulename src/cpp_source.cpp src/cuda_source.cu)

Possible solution

In pybind11/tools/pybind11Tools.cmake, line 104:
-use $<$<COMPILE_LANGUAGE:CXX>: to set the flag only for GCC
-pass it to nvcc as -Xcompiler=-flto

[YannickJadoul]

It might be even better to add this as a case to this if-block?

pybind11/tools/pybind11Tools.cmake

Line 62 in bd24155

if(CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")

Feel free to submit a PR. Otherwise, we'll put it on the backlog of "things to get to at some point".

[bstaletic]

Nvidia's docs have this section that might be interesting.

https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-guiding-compiler-driver-forward-host-compiler