fix: remove redundant deep copy in input preprocessing

[AAliKKhan]

ItemHelpers.input_to_new_input_list() already calls _to_dump_compatible() internally (items.py:772). The caller in OpenAIResponsesModel._create_request() was calling it a second time, creating an unnecessary deep copy of the entire input history on every model turn. For multi-turn conversations with large histories this doubles memory and CPU overhead.

[seratch]

Thanks for the optimization. ItemHelpers.input_to_new_input_list() already recursively materializes list inputs, so removing the second conversion is appropriate. However, test_responses_materializes_iterator_payload currently monkeypatches that helper to return an unmaterialized iterable and therefore depends on the conversion this PR removes. Please update the test to pass the nested iterable through the real ItemHelpers.input_to_new_input_list() implementation, while preserving the assertions that input and tool iterables are materialized exactly once. Once that existing regression coverage passes, this should be ready for another review.