Use Plz in OpenAI provider

[r0man]

Hi @ahyatt

This changes the OpenAI provider to use plz. This is still WIP, but I open it to get the ball rolling and so you can see how I would imagine migrating providers.

This is based on the other PR where I added the plz extensions.

Some open questions:

• I added the llm-openai-test.el file which I found useful for developing this. It calls the OpenAI API with a key at the moment, which is not ideal.

• To not test against a real server I was thinking of dumping the raw output of curl in text files for all llm-XXX generic functions for each provider. I'm not sure if it works, but in the tests we could bind the plz-curl-program to some a shell program which we generate each time we do a request. The generated program would use cat to not make a HTTP request, but instead just dump a specific text file that contains the response.

Wdyt?

[ahyatt]

I've finished the Open AI provider, and it seems to work, and the code is shorter too, so that's super nice.

The Gemini / Vertex, however, is not working, and I'm stuck on that one. As I mentioned, it uses a streaming JSON, so I don't think the current plz code can handle something like that, unless I'm mistaken. I basically need just text chunks at the minimum, but ideally it'd be the entire text so far. Let me know if it is possible and how, or if not, if you plan to add that.

[r0man]

Hi @ahyatt, nice! I will try the Open AI provider soon. I'm also going to look how I can support you with the JSON format from Gemini/Vertex.

[ahyatt]

Cool, thanks! FYI, I probably will change the Open AI provider to be closer to how you designed it initially, since I realize the Claude provider I just added uses event-streaming with completely different and much more complicated event types.

[ahyatt]

FYI, I've now changed it so that there's a special method for event-source streaming, with the logic more on the client side. I'll port Claude over as well today.