pow() with a negative base fails on non-LLVM backends

[strongoier]

When arch=ti.cpu/ti.cuda, the following script,

import taichi as ti

ti.init()

@ti.kernel
def calc_pow(i: ti.i32) -> ti.f32:
    return pow(-1.0, i)

print(calc_pow(1))

gives -1.0 , which is correct. However, it gives 1.0 when arch=ti.metal, and gives nan when arch=ti.opengl.

This is because the native pow(x, y) function in Metal Shading Language/GLSL is implemented with exp2(y * log2(x)), so it doesn't support negative bases. See https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf and https://registry.khronos.org/OpenGL/specs/gl/GLSLangSpec.4.50.pdf for more details.

The behavior is understandable, because pow() with a negative base and a real exponent results in a complex number. What we can do is to follow C++ - specially handle the case where the exponent is an integer, which results in a real number. In this way we can make the behavior of pow() consistent across all backends. I'll submit a PR for that.