PUBLIC: [tf.contrib.data] Standardize transformation functions for use with `Dataset.apply()`.

A transformation function is a function from `Dataset` to `Dataset`, and several APIs now return transformation functions, making them easier to use with `Dataset.apply()`.

This is a breaking change that modifies the semantics of `tf.contrib.data.Dataset.apply()` to remove the `args` and `kwargs` arguments; and `tf.contrib.data.group_by_window()`, `tf.contrib.data.rejection_resample()`, and `tf.contrib.data.sloppy_interleave()`, to remove the `dataset` argument. Instead the latter functions return a transformation function (from `Dataset` to `Dataset`).

PiperOrigin-RevId: 169264919