PythonDataSource is creating a repr

[vyeevani]

PythonDataSource has repr defined on it by the data class. This results in the flat examples getting serialized. This can be a huge performance penalty. The other sources have their repr's defined by the base class. Therefore, it's relatively safe to do this for tensor flow dataset as well.

[vyeevani]

This is the root cause of: google/grain#593