fix: create_schema crashes on models with a ForeignObject reverse relation

[shivanij1203]

Summary

Generating a ModelSchema / calling create_schema() for a model that is the target of a ForeignObject relation crashes with:

AttributeError: 'ForeignObjectRel' object has no attribute 'help_text'

SchemaFactory._model_fields skips reverse relations, but it only checked for ManyToOneRel and ManyToManyRel. A ForeignObject field creates a bare ForeignObjectRel (the base class of those two) on the target model, so it fell through the filter and reached field.help_text in get_schema_field, which a reverse relation does not have.

Since ManyToOneRel and ManyToManyRel both subclass ForeignObjectRel, filtering on the base class covers every reverse relation, including the bare ForeignObjectRel.

Fixes #1530

Changes

• ninja/orm/factory.py: skip ForeignObjectRel (the reverse-relation base class) instead of only its two subclasses. This also let the now-redundant cast(DjangoField, fld) be removed, since fld is narrowed to Field after the check.
• tests/test_orm_relations.py: regression test using isolate_apps so the reverse ForeignObjectRel actually registers.

Testing

• pytest tests/test_orm_relations.py tests/test_orm_metaclass.py tests/test_models.py tests/test_django_models.py passes.
• The new test fails against main with the exact AttributeError: 'ForeignObjectRel' object has no attribute 'help_text', and passes with the fix.
• ruff format --check, ruff check, and mypy ninja/orm/factory.py are clean.

Minimal reproduction (from the issue): a model OrderDetail with models.ForeignObject(Order, ..., related_name="details"), then create_schema(Order).

[mbaragiola]

Positives

• Generalizing the reverse-relation filter to ForeignObjectRel covers the bare reverse object while retaining the existing ManyToOneRel and ManyToManyRel behavior through inheritance.
• The regression test reproduces the crashing relation shape, verifies schema creation, and checks that the reverse accessor is excluded while normal fields remain.

Residual risk

• The new test covers the default depth path only, although the shared field-selection path makes a depth-specific regression unlikely.

Verification

I reviewed the relation hierarchy, schema field-selection path, and regression test against master. GitHub currently reports no checks for this head; I did not run the test suite locally.

Reviewed by Hermes Agent using Codex gpt-5.6-sol.