`shuffle_files` and `shuffle_reshuffle_each_iteration` is ignored.

Short description

shuffle_files and shuffle_reshuffle_each_iteration parameter of tfds.load() is ignored. The result dataset never get shuffled.

Environment information

• Summary: It's Google Colab
• Operating System: Linux 1803a98d124d 4.19.112+ #1 SMP Thu Jul 23 08:00:38 PDT 2020 x86_64 x86_64 x86_64 GNU/Linux
• Python version: Python 3.6.9
• tensorflow-datasets/tfds-nightly version: Reproducible with 4.0.1, 4.2.0, 4.2.0+nightly (2021-02-14)
• tensorflow/tf-nightly version: 2.4.1
• Does the issue still exists with the last tfds-nightly package?: Yes

Reproduction instructions

import tensorflow_datasets as tfds

options = tf.data.Options()
options.experimental_deterministic = False

ds = tfds.load(
    'mnist',
    split='test',
    batch_size=32,
    shuffle_files=True,
    read_config=tfds.ReadConfig(
        options=options,
        shuffle_seed=123,
        shuffle_reshuffle_each_iteration=True,
    ),
)

for entry in ds.take(2):
    print(entry['label'])
print()
for entry in ds.take(2):
    print(entry['label'])
print()
for entry in ds.take(2):
    print(entry['label'])

Actual output

tf.Tensor([2 0 4 8 7 6 0 6 3 1 8 0 7 9 8 4 5 3 4 0 6 6 3 0 2 3 6 6 7 4 9 3], shape=(32,), dtype=int64)
tf.Tensor([8 7 5 4 2 5 8 5 5 2 9 2 4 2 7 0 5 1 0 7 9 9 9 6 5 8 8 6 9 9 5 4], shape=(32,), dtype=int64)

tf.Tensor([2 0 4 8 7 6 0 6 3 1 8 0 7 9 8 4 5 3 4 0 6 6 3 0 2 3 6 6 7 4 9 3], shape=(32,), dtype=int64)
tf.Tensor([8 7 5 4 2 5 5 8 5 2 9 2 4 2 7 0 5 1 0 7 9 9 9 6 5 8 8 6 9 9 5 4], shape=(32,), dtype=int64)

tf.Tensor([2 0 4 8 7 6 0 6 3 1 8 0 7 9 8 4 5 3 4 0 6 6 3 0 2 3 6 6 7 4 9 3], shape=(32,), dtype=int64)
tf.Tensor([8 7 5 4 2 5 5 8 5 2 9 2 4 2 7 0 5 1 0 7 9 9 9 6 5 8 8 6 9 9 5 4], shape=(32,), dtype=int64)

Expected behavior

The dataset should be shuffled, and it must be reshuffled on each epoch.

Additional context

This is the 3rd day I started using Tensorflow.

Start by running the provided tfds.load() reproduction with MNIST, focusing on shuffle_files, shuffle_reshuffle_each_iteration, and ReadConfig. Done means the resulting dataset is shuffled and produces a different order on each epoch when reshuffling is enabled.