Finetune VPT model for Minetest

Finetuning would consist in a combination of
- imitation learning
- reinforcement learning

VPT is a recurrent model, which requires special treatment of the hidden state

Since VPT was developed for Minecraft it's action space has to be suitable mapped to that of Minetest.

The issue names no files, tests, or entry points. Start by locating the VPT recurrent model and the Minetest action-space mapping, then determine how imitation learning, reinforcement learning, and hidden-state handling fit together. The issue does not define a concrete completion criterion, so the scope and expected finetuning result need clarification.

Written by the indexing model from the issue text.