[Task Contribution] Meta-learning weight initialization (i.e., MAML style few-shot learning)

[51616]

I've implemented a new meta-learning task. The goal is learning the initial weight that can be finetuned in a few sgd updates efficiently. This PR is somewhat self-contain with some caveats.

• The dataset used here is omniglot imported from an external library. If needed, I can refactor and make a custom preprocessing code that can be used with raw omniglot data.
• The final accuracy for the network implemented there is not good yet. In the original MAML paper, the accuracy was near 100% for 5-way 5-shot omniglot while my network here can barely reach 80%. It could be something related to batchnorm. I can investigate more on this if the performance is a concern.

This PR serves as another good example of how one should go about implementing a new task.

Please let me know if this PR is approvable. If so, I'll make this PR review ready.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.