Wconstab/fallback infra (#71039)
* Make DivAddMul useful for fallback benchmarking
- add an aten::mm op inbetween 2 DivAddMuls
- aten::mm won't be fused (at least for now), and can be forced
to fallback, or execute as part of the lazy commputation. Since
it is never part of a fusion kernel anyway, the effect of fallback
is isolated to the fallback itself
* Add forced eager fallback
Add forced_fallback option for normal_
* Add --just_run_once to lazy_bench
* Add missing include paths
* Fix flake8