fix: detect anyOf null members by type, not exact dict equality

[nyxst4ck]

Problem

handle_null_fields removes the null member of an anyOf schema with
schema['anyOf'].remove({'type': 'null'}) — i.e. by exact dict equality.
JSON Schemas frequently annotate the null member with extra keys, e.g.
{'type': 'null', 'title': 'N'} or {'type': 'null', 'description': '...'}
(common output from MCP tools and other JSON-Schema sources). In that case no
element equals the bare {'type': 'null'}, so:

ValueError: list.remove(x): x not in list

This is reachable from the public path: t_schema → process_schema →
handle_null_fields, hit when a dict response_schema is passed to
generate_content. It also mutates the list while iterating it. Plain pydantic
emits a bare {'type': 'null'}, which is why existing tests never exercised it.

Fix

Identify null members by item.get('type') == 'null' (with an isinstance
guard) and rebuild the anyOf list instead of remove()-ing by equality. This
preserves the existing flatten-when-one-type-remains behavior, no longer mutates
during iteration, and tolerates null members with extra keys (and multiple null
members).

Testing

Added test_handle_null_fields_anyof_null_member_with_extra_keys. It raises
ValueError on main and passes with the fix. The full
tests/transformers/test_schema.py suite passes (54 passed).