Better tune the parameters in the gradient autoscheduler.

[BachiLi]

There are a few thresholds inside the gradient autoscheduler, in particular when should we use rfactor and when should we use atomics. These parameters can be better tuned on a larger dataset.