Vector reduce simplification bug.

[mcourteaux]

The extract_lanes fuzz tester found that this:

min((uint8x3)absd(x3((uint8)7), x3((uint8)3)), (uint8x3)vector_reduce_add(x9((uint8)3)))

Gets simplified to this:

x3((uint8)3

Another rabbithole bug. Gemini is trying right now.

[mcourteaux]

@abadams In general, this fuzz tester has revealed a significant amount of bugs already. I think we should let Claude make a fuzz tester much like the one we have in extract_lanes in terms of expression richness, but with more types of operations and tests.

Let's try this...

@copilot Write a fuzz tester like the extract_lanes fuzz tester in the fix-vulkan-interleave PR (#8629), but make sure to also test simplifications. We might want an optimization fence that prevents halide from simplifying anything, and pass the completely unoptimized IR to LLVM to get ground truth values to compare with after Halide simplified. One easy trick could be to add a Call intrinsic in Halide that makes the simplifier not recurse into. Afterwards, add a custom lowering pass to peel those Call nodes off, as no simplification happens after the custom lowering passes.