tests for Mullapudi2016 and Li2018 aren't very good

[steven-johnson]

The existing tests for the Mullapudi2016 and Li2018 autoschedulers aren't very useful: most of them basically just run and dump the schedule and loop nest to stdout (and maybe runtimes too), but don't attempt to validate them. Thus, they mostly amount to "can we produce a schedule without crashing and will it run in a non-crazy amount of time". This is better than nothing, I guess, but we can do better.

[abadams]

They should test that a valid schedule is generated, which is not trivial (ask any Halide beginner). That's really the only hard requirement on an autoscheduler I can think of, unfortunately. I'm not sure what else we would check.