Dataset Refactor to tfds v4 apricot issue

[shenshaw26]

Upgrading to TFDS version4 breaks the apricot datasets builder / loading using the method we are attempting in #1282 . This appears to be related to TF bug: https://github.com/tensorflow/datasets/issues/2243.

At the time of writing this ticket, we are able to build the apricot datasets however using the tfds.core.builder_from_directory method, the load fails with: ```TypeError: Only integers, slices (:), ellipsis (...), tf.newaxis (None) and scalar tf.int32/tf.int64 tensors are valid indices, got 'ragged_flat_values'

@davidslater , thoughts here?  Maybe we default back to a `tfds.load` if the `builder_from_directory` fails??

[davidslater]

builder_from_directory works but using the standard load function you can get it to work? What is load using under the hood?