Add support for cmath

[mnicely]

We should add support for cmath
https://en.cppreference.com/w/cpp/header/cmath

[wmaxey]

I would like to do this after the NVC++ rework. Currently we have the bare minimum exposed for complex.

[brycelelbach]

We'll need to coordinate with the CUDA math libraries team. I'll tentatively set it as a goal for 2.2.0, which would be early 2022.

[karthikeyann]

Usage of (present) cmath functions in Windows nvcc 11.2 causes error.
nvcc -std=c++17 cmath_test.cu --expt-relaxed-constexpr

cmath_test.cu(4): error: calling a __host__ function("exp<int, (int)0> ") from a __global__ function("kernel") is not allowed
cmath_test.cu(4): error: identifier "exp<int, (int)0> " is undefined in device code

Code:

#include <cuda/std/cmath>
__global__ void kernel() {
    cuda::std::exp(int{0});
}

int main() {
  kernel<<<1,1>>>();
}

Right now, libcudacxx cmath uses std cmath. in Linux, functions in gnu std cmath are constexpr. In MSVC cmath, it's not.

[cliffburdick]

To be more clear on the ask here, CCCL has since added support for cmath where it no longer requires --relaxed-constexpr to use the functions. However, many of the implementations in libstdc++ are inefficient on a GPU and may be better suited to either default to a path using intrinsics, or having an option to do so. Several issues have been affected by this in the past, including trigonometric functions that have multiple NaN or Inf checks that may not be needed on the GPU.