[tf.data] Add optional `shuffle` argument to `Dataset.list_files()`.

This option makes it easier to shuffle a set of filenames on each iteration,
and default to true to match the recommended best practices when training on
a large dataset.

PiperOrigin-RevId: 187434282