Support gradient_multipliers as tensor for optimize_loss

This fix tries to address the issue raised in 22295 where
gradient_multipliers for tf.contrib.layers.optimize_loss()
does not support tensor as input. This fix update the
optimize_loss to allow gradient_multipliers passed as dict
of tensors.

This fix fixes 22295.

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