Fixing finish_reason validation to allow tool_calls

[jamesbraza]

#60 added validation of the finish_reason to help paper-qa CI. However, it seems the validation is overly aggressive, as in CI we now see:

aviary.message.MalformedMessageError: Expected finish reason 'stop' in LiteLLM model response, got 'tool_calls', full response was ModelResponse(id='<redacted>', choices=[Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(function=Function(arguments='{"query":"XAI for chemical property prediction","min_year":2018,"max_year":2024}', name='paper_search'), id='<redacted>', type='function')], function_call=None))], created=1728430225, model='gpt-4o-2024-08-06', object='chat.completion', system_fingerprint='<redacted>', usage=Usage(completion_tokens=31, prompt_tokens=500, total_tokens=531, completion_tokens_details=CompletionTokensDetails(reasoning_tokens=0), prompt_tokens_details={'cached_tokens': 0}), service_tier=None) and tool choice was 'required'.

This PR thus generalizes the allowable finish_reasons when tool_choice is set to "required" (default)