add OLLAMA support

[marcnnn]

A simple Ollama Test please let me know if you like me to change things.

Thanks for this nice inspiration how to implement Agent conversations.

I would like to work on Langchain Integration as well, but I thought I start simple.

[nileshtrivedi]

@marcnnn Thanks for trying out this library! Here is how I would like to go about it:

• I would like to eliminate direct LLM calling from this repo and instead rely on https://github.com/brainlid/langchain as a unified interface for various LLMs, including for tool calling etc. To achieve this, I have been helping them test and finish the support for Gemini models.
• Therefore, this particular PR for Ollama support will soon become redundant. Let's instead work on integrating langchain v0.3.0rc which would also bring support for Ollama.
• We should stay as close as possible to Microsoft's Python-based implementation of Autogen (for eg: class names/field names etc). This will make it possible to support feature like AutoBuild where entire agent group/library can be saved to a JSON file and loaded later. There are many other tools in the Autogen ecosystem. For example, AutoGen Studio (a web app to configure and test agents), AutoGen Bench (for agent evaluations). Interoperability with Microsoft AutoGen is an important goal for us.
• I would also recommend keeping your code cleanups and refactoring separate from features/bugfixes. It keeps PRs smaller and independent which makes them easier to review and merge.
• For new features, it would be very helpful to also have a test case added.

[nileshtrivedi]

@marcnnn Since this PR changes the default LLM to LLAMA3, we need to ensure that the demo scripts continue to work.

I am okay with merging this PR if you can separate your code formatting changes from this.

[nileshtrivedi]

@marcnnn I have released v0.2.0 of autogen with Ollama as the default model. Could not merge your PR as it includes various code cleanup which conflicted with my uncommitted changes. But thanks for this contribution.

This will change soon as the next goal is to use langchain for making LLM calls in a consistent way.