fix(extensions): surface content filter refusals on the any_llm chat path

[adityasingh2400]

Some providers signal a blocked non-streaming completion only with finish_reason="content_filter" and an otherwise empty message, with no refusal field set. On the any_llm chat completions path that message converts to zero output items, so the caller sees a turn that is indistinguishable from a normal empty response and an agent loop can retry it forever.

Both siblings already handle this. OpenAIChatCompletionsModel.get_response in src/agents/models/openai_chatcompletions.py synthesizes a refusal for exactly this shape, and LitellmModel does the same for its provider-specific field. The any_llm streaming path is already covered too, because it delegates to the shared ChatCmplStreamHandler, which raises the same refusal on a terminal content_filter chunk. Only the any_llm non-streaming path was missing it, so the same provider behaves differently depending on whether you stream.

This copies the OpenAI check verbatim, including its guards: the refusal is synthesized only when the message has no content, no existing refusal, and no tool calls, so a content_filter that still carried output is left alone. Two regression tests cover both cases. The empty one fails on main with AssertionError: expected a refusal item, got: [] and passes with this change, and the non-empty one passes either way to pin the guard. make lint and make typecheck (mypy and pyright, 0 errors) are green, and tests/models/test_any_llm_model.py is 55 passed.