Use openai api instead of Ollama

[geoffroy-noel-ddh]

Currently a few operators are designed to call an Ollama api:

• answer_frames_ollama
• answer_transcription_ollama

Here we propose to send the prompt in OpenAI format instead so the operators remain compatible with Ollama but also support other engines or LLM platforms.

Advantages:

1. broader compatibility with inference engines or llm platforms
2. reduced complexity & dependencies: we can just make that call with standard Python libraries without a container
3. opportunity to move that Openai calling logic into a common ancestor operator

It also presents some difficulties as we need to preserve the following features that where supported directly (or easily) by Ollama API:

• sends a seed (for reproducible outputs)
• enable/disable thinking mode
• specify the context size in case models are loaded dynamically (e.g. Ollama)
• download the model on demand

Additional sub-tasks:

• manage errors from inference API
• make sure json results are post-processed properly
• update readme
• add param for API token

[geoffroy-noel-ddh]

Note that openai api has a standard argument to pass the seed. But thinking mode is not universal across all implementations. We are currently using the think param that ollama understands, that works well via openai api.

• But it will need adaptation and testing for other inference engines.

[geoffroy-noel-ddh]

As part of this refactoring, there's a breaking change because ollama_host parameter is renamed api_base as the intention is to be platform agnostic.