Two components of LLMcoder

[psaegert]

I've come to think that it may be best if we divide the LLMcoder copletion pipeline into two parts using two models:

1. The fine-tuned (Fine-Tuning #6) autocomplete model provides the first code suggestion as a consistently formatted and reliable baseline
2. The "feedback" model (maybe the regular gpt-3.5-turbo) continues the conversation with
   • analysis results or a "reflect"-prompt (detection)
   • a "fix this issue"-prompt (fixing)

We may need to evaluate how consistent the model is in feedback-mode, and if it suffers from the same problems regarding the format of the output and modifications of previous code (maybe we need to fine-tune another variant for the feedback-loop?)

[psaegert]

Applied in #23, #28