Fix issue with tf.nn.relu6 grad in eager

This fix tries to address the issue raised in 21380
where the gradient of tf.nn.relu6 throws out an error.
The issue seems to be that relu6 was not placed
in OpDoesntRequireInput like tf.nn.relu.

This fix fixes 21380.

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