Fix I32x4 min/max working incorrectly for large values

[Shnatsel]

The ARM I32x4::{min,max} implementations must compare integer lanes directly. Converting through f32 loses precision above 2^24, so values such as 16_777_217 round to neighboring integers and can produce a numerically wrong vector.

This PR fixes the issue and adds a regression test. The issue was spotted by GPT-5.5 xhigh.

[sagudev]

this looks related to #583 and I suspect we can also revert test changes.

This PR fixes the issue and adds a regression test. The issue was spotted by GPT-5.5 xhigh.

But the PR/code was not done by AI, right?

[Shnatsel]

The regression test/PoC is AI-written. I always have models write a PoC that fails miri before looking into the issue.

The production code change is human.

Feel free to drop the test if you have a policy against including AI contributions.