fix(parsing): prevent Pydantic schema validator leak in parse_response

[xodn348]

Summary

• Replace ParsedResponseOutputText[TextFormatT], ParsedResponseOutputMessage[TextFormatT], and ParsedResponse[TextFormatT] with their unparameterised forms in openai/lib/_parsing/_responses.py.
• Add cast(...) wrappers to preserve the static type signatures for callers.
• Add two regression tests: one for correctness, one that asserts SchemaValidator objects do not grow after the initial warm-up call.

Issue

Closes #3084

parse_response constructs ParsedResponse[TextFormatT] (and the two inner types) using a free TypeVar as the type argument. Pydantic v2's model_rebuild cannot resolve a free TypeVar, so it returns False and never populates MockCoreSchema._built_memo. This causes a fresh SchemaValidator and SchemaSerializer (heavy Rust objects) to be allocated on every single responses.parse() call. In a long-running server the process RSS grows linearly with request count.

Root cause:

# Before — free TypeVar causes pydantic model_rebuild to return False every time
construct_type_unchecked(type_=ParsedResponse[TextFormatT], value={...})

Fix — use the unparameterised class so Pydantic builds and caches the schema once:

# After
cast("ParsedResponse[TextFormatT]", construct_type_unchecked(type_=ParsedResponse, value={...}))

All three parameterised Generic fields are either guarded by if TYPE_CHECKING: (runtime-inert) or set explicitly from the dict passed to construct_type_unchecked, so the unparameterised form is runtime-equivalent.

Local verification

=== LOCAL_TEST_PASSED ===
$ cd /tmp/openai-python && python3 -m pytest tests/lib/responses/ tests/test_models.py tests/test_utils/ -v --override-ini="addopts=" -q

============================= test session starts ==============================
platform linux -- Python 3.11.15, pytest-9.0.3, pluggy-1.6.0
rootdir: /tmp/openai-python
configfile: pyproject.toml
plugins: respx-0.23.1, inline-snapshot-0.33.0, xdist-3.8.0, asyncio-1.3.0, anyio-4.46.0
asyncio: mode=Mode.AUTO

collected 186 items

tests/lib/responses/test_responses.py .......                            [  3%]
tests/test_models.py ................................................    [ 29%]
tests/test_utils/test_datetime_parse.py ...................................
...............................                                          [ 63%]
tests/test_utils/test_json.py .........                                  [ 68%]
tests/test_utils/test_logging.py .....                                   [ 70%]
tests/test_utils/test_path.py ..............................................
.....                                                                    [ 96%]
tests/test_utils/test_proxy.py ..                                        [ 97%]
tests/test_utils/test_typing.py .....                                    [100%]

186 passed in 0.86s

Schema-leak verification (warm-up + 50 calls, delta must be 0):

gc.collect(); before = count_schema_validators()
for _ in range(50): parse_response(text_format=CalendarEvent, ...)
gc.collect(); after = count_schema_validators()
# delta = 0  ✓

Risk

• Correctness: no runtime change — the type arguments only affect type-checker annotations. The parsed field value is already embedded in the dict passed to construct_type_unchecked.
• Type safety: explicit cast preserves all existing type signatures; no downstream callers need updating.
• Pydantic v1 / v2: both use construct_type_unchecked the same way; the caching fix only matters for Pydantic v2's model_rebuild path, and the unparameterised form is valid for both.