Incorrect cos/sin on arrays with more than 2 elements with ffast-math

[gftabor]

Minimal reproducible example. https://github.com/gftabor/xtensor-test

#include <iostream>
#include <xtensor/xarray.hpp>
#include <xtensor/xio.hpp>
#include <xtensor/xview.hpp>

int main(int argc, char* argv[])
{
    xt::xarray<double> array = xt::eval(xt::zeros<double>({2,2}));
    auto res_incorrect = xt::eval(xt::cos(array));
    std::cout << res_incorrect << std::endl;
    xt::xarray<double> small_array = xt::eval(xt::zeros<double>({2}));
    auto res_correct = xt::eval(xt::cos(small_array));
    std::cout << res_correct << std::endl;
    return 0;
}

(mqp) tabor@taborwork:~/Documents/mqp/xtensor-demo/build$ ./xtest 
{{-1., -1.},
 {-1., -1.}}
{ 1.,  1.}

Specifically this is with xtensor 0.25 and xsimd 11.2. It works correctly with xtensor 0.25 xsimd 12.1.1 .