We always tend to install the optionals for testing but it
turns out the import tests were not working properly.

Consider removing S3 completely.

Somehow pydantic v1 did not notice.

Still some failures. Lots of deprecation warnings.

AstropyArrow and Dataframe use classes that are not part of
the core butler dependencies. This leads to the tests failing
in a minimal environment. Replace with two storage classes
that are available to the core environment.

This enables pydantic v2

This is meant to be the rquivalent of using __setattr__
and __field_set__ directly. The latter has been renamed in
pydantic v2

This all worked before but now we have to force the Config
to a plain dict before it will be accepted.

There is clearly some issue with the RootModel constructor
that we will need to sort out.

This required that we allow unused igores since the location
of the ignores depends on the version of pydantic and mypy
always picks up the first definition it sees even if that
is a pydantic V2 code path and pydantic is v1.

If the BaseModel.construct() method is slow we can still
reimplement it as was done in the individual direct() methods
but with the direct() methods still being much simpler.

The previous version used by the direct() methods
would be something like:

        def model_construct(cls, _fields_set: set[str] | None = None, **values: Any) -> Self:
            node = cls.__new__(cls)
            setter = object.__setattr__
            for k, v in values.items():
                setter(node, k, v)
            if _fields_set is None:
                _fields_set = set(values.keys())
            setter(node, "__fields_set__", _fields_set)
            return node

Testing with 100MB graph indicates that construct() might be
fractionally slower than implementing it ourselves as above
but it is not conclusive. The direct() method only accounts
for 14% of the graph load time.