Fix discrepancies between doc and implementation for math_ops

In the docstring of math_ops, there are some discrepancies
between the supported data types and the actual implementation:
`abs` supports `float16`.

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