Propagate some facts about inequalities with min/max

[shoaibkamil]

Fixes #8443.

Teaches the simplifier to propagate some facts about inequalities where one of the sides is a min or max node.

[mcourteaux]

Are the inverse statements required?

[abadams]

Same question. Is this just to get a syntactic match on the <= form? The other cases here don't do it, but I think this is not exactly the same as the other cases.

[shoaibkamil]

I found it was necessary i.e. if I comment out the learn_false() lines, the test fails, since we can't eliminate the condition in the third specialization. I added all the learn_false() clauses even though it may be true that only one or two are needed for this testcase.

We have a rewrite that changes downsample = !(scale_factor_x > 1.0) && !(scale_factor_y > 1.0) into downsample = max(scale_factor_x, scale_factor_y) <= 1.0. Without these learn_false() clauses, the simplifier is unable to prove max(scale_factor_x, scale_factor_y) <= 1.0 --> !(1.0 < scale_factor_x) and similarly max(scale_factor_x, scale_factor_y) <= 1.0 --> !(1.0 < scale_factor_y), so it can't trim the conditions in that branch.

[mcourteaux]

While this PR might fix the issue in the particular case of #8443, there is more that needs to be done, according to @abadams comment:

specializations should be resolved before the simplifier gets a turn at it.

[shoaibkamil]

Is the failure for performance_inner_loop_parallel a fluke or something I should look into? Otherwise, failures seem to be in vulkan, and one for the autoscheduler test.

[abadams]

I spot-checked that test locally and it seemed fine, so I'm hoping it was a fluke.

[shoaibkamil]

I also verified locally that the IR for that test did not change after this change.