Remove `non-fused` version of `adjust_saturation` as GPU kernel already exists (#14794)

* Remove `non-fused` version of `adjust_saturation` as GPU kernel already exists

In the existing implementation for `adjust_saturation` the non-fused version
was still in place. As the non-fused is for non-GPU support of `adjust_saturation`
and GPU kernel already exists now (See commit https://github.com/tensorflow/tensorflow/commit/25c4f279402c3f9516314ee40ee398fc57ffb78e#diff-b53c223158b7c4fd248ef581da6566c2

), it makes sense to remove the non-fused version.

In addition, with the removal of non-fused implementation of `adjust_saturation`,
now it is possible to provide batch support (in 4-D instead of previous 3-D).
This resolves issue raised in 8926.

This fix removed non-fused version of `adjust_saturation` and added additional
test cases for batch support.

Note: In PR 14187, non-fused version of `adjust_hue` has been removed so
batch support for `adjust_hue` has been enabled as well. This PR also
adds additional test cases for batch support of `adjust_hue`.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add batch support for `adjust_saturation` and ``adjust_hue``

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Update docstring for `random_hue`

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>