Atmoic min get wrong result on cuda backend when block_dim is 16

[BillXu2000]

Describe the bug
Atmoic min get wrong result on cuda backend when block_dim is 16.

To Reproduce
Run the code below. The expected answer is 1, but I got 0 on my computer.

import taichi as ti

ti.init(arch=ti.cuda)

n = 16
x = ti.field(dtype=ti.f32, shape=n)

@ti.kernel
def test():
    hit_time = 1
    for i in x:
        ti.atomic_min(hit_time, 1)
    print(hit_time)

test()

Setting n to any number larger than 16, print value of hit_time, or replacing for i in x with for i in range(n) will get the right answer.
Especially, replacing for i in x with for i in range(n) will only change the block_dim in the final Taichi IR, and thus I think the bug might be at cuda codegen phase.

[BillXu2000]

Related (same) issue: #5142

[bobcao3]

Can you try to disable reduction optimization and see whether that fixes this?

[BillXu2000]

Disabling reduction optimization will fix this bug.