Registered functions with pydantic BaseModel arguments are returned as dict

[NixBiks]

I'm not sure whether this is intended behavior or not. In the following example I'd expect make_resource function to receive a Resource.

from typing import Callable
from pydantic.main import BaseModel
import thinc

from thinc.api import Config, registry


class Resource(BaseModel):

    url: str


thinc.registry.create("resources")


@thinc.registry.resources("Resource.v1")
def make_resource(
    resource: Resource,
) -> str:
    return resource.url  # this fails since resource is a dict and not Resource


config = Config().from_str(
    """
[test]
@resources = "Resource.v1"

[test.resource]
url = "https://thinc.ai/docs/usage-config"
"""
)
resolved = registry.resolve(config)

But the current registry.resolve simply validates that a dict would be able to be parsed to Resource but is not actually resolved to a Resource, i.e. the type in make_resource is a dict in the example above so I end up with the following error

AttributeError: 'dict' object has no attribute 'url'

Now obviously the validation should pass but I think the dict should be casted into the given BaseModel as well so the typings in the registered functions can be trusted.

I imagine this is related to #57 and #59