Under mock_data, a trivial decoder changes the output tensor

Short description
When tfds.testing.mock_data is active, passing a custom decoder to tfds.load alters the result, even if the decoder doesn't actually change the decoding logic.

Environment information

• Operating System: Ubuntu 18.04

• Python version: 3.7.5

• tensorflow-datasets/tfds-nightly version: tensorflow-datasets 4.5.2

• tensorflow/tf-nightly version: tensorflow 2.7.0

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

Reproduction instructions

import contextlib
import tensorflow as tf
import tensorflow_datasets as tfds

@tfds.decode.make_decoder()
def no_op_decoder(example, feature):
    return feature.decode_example(example)

with tfds.testing.mock_data():
#with contextlib.nullcontext():
    ds = tfds.load('imagenet_v2', split='test')
    image1 = next(iter(ds))['image']

    ds = tfds.load('imagenet_v2', split='test',
        decoders={'image': no_op_decoder()})
    image2 = next(iter(ds))['image']

    print(tf.reduce_all(image1 == image2).numpy())

Running this prints False, but if you comment the mock_data line and uncomment the nullcontext line, it prints True.

Link to logs
N/A

Expected behavior
It should print True in both cases.

Additional context
N/A

Start with the tfds.testing.mock_data context, tfds.load, and the tfds.decode.make_decoder entry points shown in the reproduction. Compare the two loading paths and add or run a regression test using the no-op decoder; done means both paths produce equal image tensors, as the example expects.