bug: serializing and then deserializing to/from json or dict does not preserve dataclass properties

[JohannesMessner]

How to reproduce:

from docarray import Document, dataclass
from docarray.typing import Text


@dataclass
class MyMMDoc:
    t: Text
    

m = MyMMDoc(t='hi')
d = Document(m)
d.is_multimodal
Out[0]: True
d_proto = Document.from_protobuf(d.to_protobuf())
d_proto.is_multimodal
Out[1]: True
d_dict = Document.from_dict(d.to_dict())
d_dict.is_multimodal
Out[2]: False
d_json = Document.from_json(d.to_json())
d_json.is_multimodal
Out[2]: False

Probable Cause:

When calling d.to_dict() or d.to_json(), d gets converted through to_pydantic_model() to a PydanticDocument.
The PydanticDocument contains _metadata (which is responsible for enabling multimodal features) as a private attribute, and private attributes are not kept by pydantic (more info (here)[https://pydantic-docs.helpmanual.io/usage/models/#private-model-attributes]).

Possible Solution:

One idea to solve this coild be with a small special logic that converts _metadata to metadata when converting to PydanticDocument.

Something like this:

    def to_pydantic_model(self) -> 'PydanticDocument':
        """Convert a Document object into a Pydantic model."""
        from ..pydantic_model import PydanticDocument as DP

        _p_dict = {}
        for f in self.non_empty_fields:
            v = getattr(self, f)
            ...
            elif f == '_metadata':
                _p_dict['metadata'] = v
            else:
                _p_dict[f] = v
        return DP(**_p_dict)

And the equivalent in from_pydantic_model().