Modify the openai model to conform to the new openai sdk v1.0.0

[RobinPicard]

This PR aims at solving this issue: #361

The new Python sdk of Openai allows us to instantiate a client that we can then use for API calls. I thought it would make sense to put the instantiation of the client (and the import of the openai library) in the __init__ of the OpenAIAPI class.

An issue I ran into was that the functions call_chat_completion_api and call_completion_api would then receive the client as an argument which would cause problems with the cache decorator. I'm not sure what's the best way to handle that. I created two intermediary functions cached_call_completion_api and cached_call_chat_completion_api in the __init__ of the OpenAIAPI class as it looked like the solution that required the fewest modifications to the existing logic.

The new sdk handles the retry logic itself so I removed the retry decorator (and added a max_retries argument to OpenAIAPI instead)

[rlouf]

This works locally, thank you so much for this great contribution!