Erroneous f32 -> i32 casting on the CUDA backend

[lucifer1004]

Describe the bug
f32 -> i32 casting yields erroneous results for certain inputs on the CUDA backend.

To Reproduce

import taichi as ti

ti.init(arch=ti.cuda)
x = ti.field(dtype=ti.f32, shape=(1,))

@ti.kernel
def test(T:ti.f32):
    x[0] = int(abs(T) / T)

errors = []
for t in range(1, 1025):
    test(t)
    if x[0] == 0:
        errors.append(t)

print(len(errors))
print(errors)

Log/Screenshots

[Taichi] version 1.7.0, llvm 16.0.0git, commit 2fd24490, linux, python 3.11.8
[Taichi] Starting on arch=cuda
104
[33, 37, 41, 47, 55, 66, 74, 75, 82, 94, 107, 109, 110, 115, 121, 132, 139, 148, 150, 159, 164, 188, 214, 218, 220, 227, 230, 235, 242, 245, 255, 264, 278, 296, 300, 318, 328, 376, 427, 428, 436, 440, 454, 460, 470, 477, 484, 490, 495, 497, 510, 515, 519, 528, 535, 543, 547, 555, 556, 567, 575, 592, 600, 619, 636, 647, 649, 656, 691, 749, 751, 752, 787, 791, 803, 819, 833, 854, 856, 867, 872, 879, 880, 887, 899, 908, 917, 920, 927, 933, 940, 953, 954, 968, 975, 980, 981, 987, 989, 990, 994, 1011, 1017, 1020]