Go over examples

[abheesht17]

There are a few things we need to look into for examples.

• For the GRU4Rec example, we need to figure out if no loss reduction is the right way to do it (instead of sum_over_batch_size). The loss does not converge for the latter, whereas it does for the former. TFRS does no reduction.
• Is the TFRS way of splitting sequences randomly to get train-test set correct? Won't there be data contamination this way?
• Consider shifting back to a subclassed model for DCN (rather than Functional).
• Data parallel example: curious case of giving the same prediction for every user. Needs to be fixed.