Update Qwen state training and inference code

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

Upload ATCTrack-VLM code and selected checkpoints