[Idea] Compile relay -> pytorch/tensroflow eager, and remove all pytorch/tf direct benchmark.

[MarisaKirisame]

No description provided.

[slyubomirsky]

I would argue the other way around (using Relay to import models from other frameworks) might be more helpful for experiments that we want to use in papers: We would want to convince reviewers that our system (which they probably will not know as well) is doing the same thing as established ones

[MarisaKirisame]

compiling python is hard. but either is fine.