Support pydantic dataclasses

[eric-czech]

I have had no issues using Pydantic dataclasses with draccus so far except that field descriptions are not parsed correctly in CLI help strings, e.g.:

# > cat pydantic_example.py
from pydantic import Field
from pydantic.dataclasses import dataclass
import draccus


@dataclass
class Config:

    batch_size: int = Field(
        default=128, 
        description="Batch size for training"
    )

@draccus.wrap()
def main(cfg: Config):
    """Main function."""
    print(cfg)


if __name__ == "__main__":
    main()

Show CLI help:

python pydantic_example.py --help
...
Config:

  --batch_size int    #  <---- The field description is ignored

Similarly, adding a comment near the field definition is both redundant and demonstrates that default values aren't processed correctly (i.e. the Field wrapper is not unwrapped):

+   # Batch size for training
    batch_size: int = Field(
        default=128, 
        description="Batch size for training"
    )

Show the help again:

python pydantic_example.py --help
...
Config:

  --batch_size int   Batch size for training (default: annotation=NoneType required=False default=128
                     description='Batch size for training')

Now there is a help comment, but not a great one.

---

I'll also note that It may be possible to support metadata in standard dataclass fields with a similar solution, e.g. from https://docs.pydantic.dev/latest/concepts/dataclasses/:

The dataclasses.field used in this case comes from the Python standard library: https://docs.python.org/3/library/dataclasses.html#dataclasses.field.

Perhaps looking for both title and metadata in either dataclass field metadata or Pydantic Field instances would be a reasonable start.

Another better option IMO would be to provide an API allowing users to determine how necessary field metadata is parsed. It looks like draccus/wrappers/docstring.py#L65 would be where such a thing would need to plug in?