`tf.GradientTape` inside a `tf.function`.
    *   Changed the default step size in `gradient_checker_v2.compute_gradients`
        to be exactly representable as a binary floating point numbers. This
        avoids poluting gradient approximations needlessly, which is some cases
        leads to false negatives in op gradient tests.
    *   Added `tf.config.experimental.get_memory_info`, returning a dict with