eager: Update use of Datasets.

1. Use tf.data instead of tf.contrib.data
2. Move the CPU->device copy to tfe.Iterator.
   (So that users of tfe.Iterator don't have to do so themselves)

The latter is implemented using tf.identity() to explicitly copy
each element of the dataset, which can become a bottleneck. This
should be replaced by a scheme there the dataset prefetches elements
into device memory.

PiperOrigin-RevId: 173607457