Hopefully fix flaky mullapudi reorder test

[abadams]

This tested vectorized code against scalar code, but the scalar code was something that could be potentially be optimized by llvm quite cleverly - it was just a linear combination of the loop variables. This changes the test to be less of a weird corner case, while still preserving the important property - the evaluation of f is fine under any storage ordering. Hopefully this will make it flake less often.

[alexreinking]

Still flaking on arm-64-osx:

Autoscheduler time (1) is slower than expected:
======================
Manual time: 10.5701ms
Auto time: 22.7106ms
======================

https://buildbot.halide-lang.org/master/#/builders/171/builds/350

[abadams]

Sigh. Not as badly at least. Previously the error threshold was set to 15x slower. I set it to 2x slower after fixing it, but maybe that wasn't conservative enough.