[WordSeg] Add example

[texasmichelle]

No description provided.

[asuhan]

Thanks!

[BradLarson]

Looks great (assuming the test failure is flaky and can be resolved). We should make a note of the NaN testing case as a future callback we need for the unified training loop. Keras has this as a callback, and I can see other models using that.

[texasmichelle]

The test failure is indeed puzzling. My hunch is that kokoro was using a cached image, so I'm curious to see whether it passes without any changes to TextUnsupervised.

Do we have an open issue for training loop creation? Would that be here or in another repo?

[texasmichelle]

I couldn't get past the first epoch with the full dataset, so I added support for a smaller dataset, which trains quickly. The full dataset gives us our performance target and the flexibility to use a smaller sample will help get us there.