Working on a more human-like LLM player harness in the `llm_agent_loop_nodejs` file.

Bugfix: incorrect force array computation due to concatenation of `valid_mask` on multihot level (which in general was a bad idea and should be fixed)

Implement `flickscreen` in NodeJS and CPP renderers.

Some minor JAX optimizations and benchmarking scripts.

Work on training diff model.

Some work on prompting LLM player with past moves.

A bunch of work e.g. on batched nodejs

Add missing profile_rand_cpp script

Work on agent loop with local vllm-hosted model.

Bugfix CPP PuzzleScript

Work on CPP PuzzleScript version

Remove deprecated random rollout profiling logic from `search_nodejs.py`

some missing JS backend stuff

Tuck away deprecated hack of original PuzzleScript repo in `script_doctor` where it was used. NodeJS-native random rollout workers for profiling.

Adding nodeJS backend to profiling (with multiprocessing), also `evolve_level_nodejs` and `train_nodejs` (only single-env, dense for now