Add complex dtypes support for `tf.squared_difference`

This fix tries to address the issue raised in 14932 where
complex dtypes are not supported for `tf.squared_difference`,
which is different from the doc string in `math_ops.cc`
(see `BINARY_FEWER`).

This fix adds the complex64 and complex128 support in kernel,
and adds additional test cases.

This fix fixes 14932.

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