Fix decoding error on assistants response

[lzell]

The response_format and tool_choice fields of an assistant's run can either be objects or strings.

Previously, if the tool_choice was a string (e.g. 'auto') in the run response, we failed to decode the response. Similarly, if the responses_type was a string (e.g. 'auto') in the run response, we failed to decode the response.

This bug was reported by @mbrian23, with a full repro here. Thank you Martin!
https://gist.github.com/mbrian23/6863ffa705ccbb5097bd07efb2355a30

I added regression tests for the failing case.

References:
https://platform.openai.com/docs/api-reference/runs/createRun#runs-createrun-response_format
https://platform.openai.com/docs/api-reference/runs/createRun#runs-createrun-tool_choice

[jamesrochabrun]

Thanks so much @lzell ! Also huge thanks dding tests!