Fix: retry on invalid JSON from OpenAI and Claude

[runarmod]

This PR adds logic to retry prompting OpenAI and Claude if their JSON response is not valid. See #829

Note: The implementation of the JSON validation for Claude is different than the others, so I had to use an alternative solution. It looks like it is possible to instruct Claude to produce JSON based on a schema, as is done for the other services. I haven't used marker with the Claude service, so I don't know if there are many JSON mistakes when not using a schema. It might be worth looking into for someone using Claude with marker. See the documentation.

[runarmod]

Latest commit reverts sleeping after catching JSON parsing error, as was done with Gemini in #848.

[VikParuchuri]

Thank you! I updated the gemini service to also adjust the temperature if json failed to be parsed, so we avoid getting the same result again - https://github.com/datalab-to/marker/blob/master/marker/services/gemini.py . Is it possible to do something similar here?

Re: claude, I think when I implemented this originally, claude didn't have support for structured output. Will refactor at some point, thanks for the pointer.

[runarmod]

Thank you! I updated the gemini service to also adjust the temperature if json failed to be parsed, so we avoid getting the same result again - https://github.com/datalab-to/marker/blob/master/marker/services/gemini.py . Is it possible to do something similar here?

I see that temperature control is easily modifiable via the parse/create calls for these services as well, so I can implement that here as well.

What do you think is a good baseline temperature for the other services? I see that the default temperature specified (by marker) for Gemini is 0.0 (default specified by Google depends on the model, but possible range is 0.0-2.0)(source). The other services do not have a temperature specified (by marker), so they use the defults defined by Anthropic and OpenAI.

I see that Anthropic has a default of 1.0 (possible range 0.0-1.0) (source), and OpenAI defaults to 1.0 (with possible range 0.0-2.0) (source).

Should we hardcode a default of 0.0 instead, and then increase to 0.2 if the JSON is malformed? Or 0.1 for Anthropic since they use a smaller range?