[Relay][Prelude] Remove Peano nats from the prelude

[slyubomirsky]

@jroesch and others have noted that the Peano (unary) nat, other than being a convenient example for ADTs, is not actually very useful outside of proof assistants and should be replaced with integer scalars.

This PR removes nats from the prelude and replaces them with integer scalars in other functions. Unfortunately, there appear to be issues with the interpreter that prevent the scalar operations from properly being lowered, so I have left this PR as a WIP until those issues can be fixed and allow the tests to work. (@MarisaKirisame says the bug encountered here is most likely fixed in the VM PR.)

[slyubomirsky]

Please review @jroesch, @MarisaKirisame, @zhiics

[slyubomirsky]

Indeed, the VM updates have unblocked this PR, so I think it can be reviewed now.

@MarisaKirisame I think test_pow() in test_pass_gradient.py is failing because of these changes but I don't know what could be causing it. Could you have a look at it and see what the issue might be?
edit: The gradient fails to type check but I do not know what the error is; seemsl ike bad error handling there

I think it may be a problem with the actual gradient produced because the type checker reports failing to unify (Tensor[...], Ref[...]) with Tensor[...]

edit: I think the error is because DeGlobal in gradient.cc can't handle recursion inside a global definition (the global vars are left in the gradient), so I will change the test and I think this error should be fixed in a separate PR.

[MarisaKirisame]

it is failing because test_pow use iterate which use natrual number. it is design specifically to test adt and pattern matching work for reverse mode. can you add nat in that test?

[slyubomirsky]

I think the type error is happening because there are globals (namely iterate) left in the final gradient, even though the gradient pass assumes that DeGlobal gets rid of them all (but it doesn't handle global functions that are recursive)

[tqchen]

ping @MarisaKirisame @jroesch @joshpoll please review again and see if we can move on to merge it in

[jroesch]

@slyubomirsky @MarisaKirisame did you resolve the previous thread of conversation?

[MarisaKirisame]

@slyubomirsky can you move the nats into a common file, and restore the test in anf and gradient into exactly the same as before? I am explicitly trying to test pattern matching there.

[slyubomirsky]

That will be an easy change, thanks for explaining those tests for me.

[tqchen]

Thanks, @MarisaKirisame @slyubomirsky @jroesch , this is now merged