[tf.data] Adding support for `tf.data.AUTOTUNE` as a special value for the...

[tf.data] Adding support for `tf.data.AUTOTUNE` as a special value for the `num_parallel_calls` argument of `tf.data.Dataset.map()`, `tf.data.Dataset.interleave()`, and `tf.contrib.data.map_and_batch()`.

When `tf.data.AUTOTUNE` is specified, the level of parallelism is determined at runtime. The underlying mechanism instruments the input pipeline to build a performance model and then uses the model to find the optimal values for the parallelism knobs.

PiperOrigin-RevId: 213283297