Add `double` support for `tf.decode_csv`

In the current tensorflow `tf.decode_csv` accepts
`float`, `int32`, `int64`, `string` but not `double`.
It seems adding `double` support makes sense as `StringToNumber`
already support `double` type.

This fix adds `double` support for `tf.decode_csv`

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