Fixes #1263

When a union had a const member that wasn't the last option, the generated parser ran the const check with no type check or try/except around it, so it returned (or raised) before trying the other variants. That shows up in a few places:

• Responses like anyOf: [{const: "x"}, {$ref: Model}] (the spec in anyOf in responses generates invalid body parser #1263): the model branch is unreachable, so a dict response comes back as the raw dict.
• Model properties like anyOf: [{const: "auto"}, {type: integer}]: from_dict({"sizeOrAuto": 5}) raises ValueError.
• prefixItems with a const prefix (already in the 3.1 golden record): any other string in the array raised ValueError.
• to_multipart() on a model with a const inside a union generated isinstance(x, Literal["auto"]), which raises TypeError: Subscripted generics cannot be used with class and instance checks.

The change gives ConstProperty a get_instance_type_string() (the runtime type of the const value: str, int, float or bool) and a check_type_for_construct macro, so consts take the same try / type check / fallthrough path in union_property.py.jinja as every other member, and the multipart isinstance check uses a real type.

One behavior note: for oneOf: [{type: "null"}, {const: "x"}], a non-matching string now falls through to the final cast(...) like other union members already do (e.g. oneOf: [null, date]), instead of raising. The two golden-record diffs show exactly that.

Tests:

• New functional tests in test_unions.py for model properties (const first, mixed str/int consts with a boolean, const plus $ref), a response union using a mocked httpx client, and to_multipart(). All but the type-hint test fail on main.
• Unit test for ConstProperty.get_instance_type_string().
• Regenerated golden records, then ruff check ., ruff format --check ., mypy openapi_python_client fuzz, and pytest --cov openapi_python_client tests end_to_end_tests (475 passed, 100% coverage).