Clean up docs/ and issues/ directories from main branch

Merge pull request #43 from tokenbender/full-vocab-softmax-gradient-flow

config: adjust training parameters for full vocab experiments

feat: compute softmax over full vocabulary for complete gradient flow

refactor: move hardcoded values to config for better flexibility

add optimizations with vectorized operations, action space based probs calculation, removal of entropy calculation over entite vocab

Merge branch 'optimize-avatarl-vectorization'

add basic timing bash file and corresponding small iteration config

Update training configuration and fix evaluation timing

fix critic download path reading from config file in start.sh

Merge pull request #41 from tokenbender/feat/4bit-distributed-critic

feat: add 4-bit quantization support for critic model in AvataRL

feat: add 4-bit quantization support to sample.py

perf: optimize checkpoint loading to reduce memory usage

feat: add 4-bit quantized critic loading for memory optimization