v0.3.0: minimal prototype edition / switch to litellm and uv

The overall goal was to create a minimal understandable prototype that is well-suited to be used as base for new experiments as well as a baseline for other projects. Focusing on simplicity allows both humans and LLMs to easily use and adopt the code.

Changes include:

• switch from langchain to litellm to reduce dependencies and make all LLM interactions as transparent as possible.
• use function-calling/tool-calling for all LLM agent interactions and external tool integration
• move from single-prompt LLM-interactions to providing full message histories to the LLM
• switch to uv for dependency management
• add time-limiter to stop test-runs/samples after x seconds; this is best-effort, not a hard limit.
• context-management: add a configurable context-size limit for the high-level strategy loop. When the limit is reached, the high-level LLM is asked to compact/compress history.
• updated JSON log format and adapted all analysis tools and replayer to this format
• removed wintermute to create a minimal prototype

Somewhat fitting, today is Easter Night. In Faust, this is the time when Faust has reached despair, convinced that all his scholarship has brought him no closer to understanding the world.