Cleanup: remove checkpoint file

Cleanup: remove tensorboard/

Cleanup: remove output/

Cleanup: remove scripts/train_qwen_state.sh

Update: two-stage training, per-channel FiLM gate, cosine scheduler, 9B config

Update Qwen target-state forward teacher training docs

Update wandb logging and training docs

Upload ATCTrack-VLM code and selected checkpoints

initial commit