Better parsing of Numpy docstring with empty section indent
Also field parsing is stricter now and it can't be parsed if parameter
name isn't valid Python identifier.
As soon as I fixed parsing of Numpy docstring format it caused errors in
multiple tests that used types of function parameters, because
previously in these places docstrings couldn't have been parsed
successfully (they were treated as Epydoc docstrings) and
PyNamedParameterImpl#getType delegated to NumpyDocStringTypeProvider.
I explicitly prohibited using of NumpyDocString for this purpose before
NumpyDocStringTypeProvider is migrated to the newer docstring API.