Incorrect results are produced when using the pow function with float16 arguments with the CUDA backend

Description

The following definition for the pow function is used in the JIT compiled kernel when the arguments are float16:

#ifdef AF_WITH_FAST_MATH
#define __pow(lhs, rhs)  \\
static_cast<double>( \\
pow(static_cast<double>(lhs), static_cast<double>(rhs)));
#else
#define __pow(lhs, rhs) \\
__float2int_rn(powf(__int2float_rn((int)lhs), __int2float_rn((int)rhs)))
#endif

As a result, if AF_WITH_FAST_MATH is false, the arguments are cast as integers and the result is rounded to the nearest integer.

Reproducible Code and/or Steps

#include <arrayfire.h>

using namespace af;

int main(int, char **) {
        info();

        array A = randu(64, 64, f16);
        array B = randu(64, 64, f16);
        array A_32(64, 64, f32);
        A_32(span,span) = A(span,span);
        array B_32(64, 64, f32);
        B_32(span,span) = B(span,span);
        array E = pow(A, B);
        array E_32 = pow(A_32, B_32);

        array E_d = abs((E - E_32) / E_32);

        float tol = 1.0e-3;
        float max_d;
        unsigned loc_d;
        max(&max_d, &loc_d, E_d);
        if(max_d > tol) {
            throw;
        }

    return 0;
}

Expected result: max_d <= 0.001 when comparing half precision with float
Actual result: max_d > 0.001

System Information

ArrayFire v3.9.0 (CUDA, 64-bit Linux, build 5e1e336)
Platform: CUDA Runtime 12.4, Driver: 550.54.15

Distributor ID: Ubuntu
Description: Ubuntu 22.04.4 LTS
Release: 22.04
Codename: jammy
name, memory.total [MiB], driver_version
NVIDIA GeForce RTX 4060 Laptop GPU, 8188 MiB, 550.54.15

Checklist

• [ x] Using the latest available ArrayFire release
• [ x] GPU drivers are up to date