mock_data doesn't work with VOC when custom decoders are used

Short description
Attempting to iterate over the dataset returned by tfds.load('voc', decoders=...) when tfds.testing.mock_data is in effect results in an exception.

I assume this affects pretty much any dataset with a feature of type tfds.features.Sequence.

Environment information

• Operating System: Ubuntu 18.04

• Python version: 3.6.9

• tensorflow-datasets/tfds-nightly version: tensorflow-datasets 4.4.0

• tensorflow/tf-nightly version: tensorflow 2.6.2

• Does the issue still exists with the last tfds-nightly package (pip install --upgrade tfds-nightly) ? Yes

Reproduction instructions

import tensorflow_datasets as tfds

with tfds.testing.mock_data():
    voc = tfds.load('voc', split='test', decoders={
        'image': tfds.decode.SkipDecoding(),
    })
    print(next(iter(voc)))

Link to logs
https://gist.github.com/IRDonch/7275316c5a1323c9f4a22d002b8391d9

Expected behavior
Valid examples should be returned.

Additional context
N/A

Start by running the provided tfds.load and tfds.testing.mock_data reproduction with the VOC dataset, then trace how custom decoders interact with tfds.features.Sequence. Done means iteration returns valid examples without an exception, with coverage for the reported decoder configuration.