Add test cases for int64 support of `tf.reduce_max` on GPU

This commit is part of the effort to support int64 for
`tf.reduce_max` on GPU.

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