Remove incorrect MAML implementation

[zou3519]

Fixes #328

Here's some context:

• We discovered that the implementation doesn't actually use maml (Does the hessian really gets computed? dragen1860/MAML-Pytorch#59)
• We filed Replace MAML with an actually correct implementation #328
• We added maml_omniglot Add example of maml training on omniglot #349 as the correct version of of the maml model.
• We didn't delete the maml model (because I was worried that it was doing a "different" type of maml that I hadn't seen before that is still valid).

The last step to resolve this issue is to delete the incorrect MAML example, unless we have reasons to keep it around.

[xuzhao9]

It makes sense to me to remove maml and keep maml_omniglot as the correct impl.
Another stakeholder of the maml model is the torchdynamo team, @jansel are you okay with this PR?

[jansel]

maml is one of the hardest models for dynamo to handle because it involves a copy.deepcopy() and some dynamic autograd during inference.

I think it is pretty useful and would like to keep it, though perhaps more of a challenging test case and less as a benchmark.

[xuzhao9]

I see, let's keep this model then. I will close this PR, and create a separate PR to mark maml as correctness test only and will not be used for performance testing.

Thanks @jansel and @zou3519 !