Support optional tensors

TF recently added optional support: https://www.tensorflow.org/api_docs/python/tf/experimental/Optional
It seems to be working inside tf.data:

ds = tf.data.Dataset.range(3)
ds = ds.map(lambda x: tf.experimental.Optional.from_value(x))
list(ds)[0].get_value()

We should update our feature connectors to accept None to return optionals. Maybe with a tfds.features.Optional like

def _info(self):
  features=FeatureDict({
      'image': tfds.features.Optional(tfds.features.Image()),
  })

def _generate_examples(self):
  yield {
      'image': None,  # Accept `None` or feature
  }

Which would returns:

ds = tfds.load()
ds.element_spec = {'image': OptionalSpec(TensorSpec(shape=(None, None, 3), dtype=tf.uint8, name=None),)}

Not clear how this would translate into tf.train.Example spec though.

Start by reviewing TFDS feature connectors and the TensorFlow Optional API, then trace how feature values are represented in tf.train.Example. Determine the intended Optional connector behavior for None, including element_spec and serialization, and add coverage for accepted optional values once the representation is decided.