SPaTS training configs

We do not release SPaTS checkpoints. Reproduce from public Qwen-VL base models:

Stage	Config	Starts from	Saves to
1. SFT	configs/sft/sft_2b.yaml / sft_4b.yaml	Qwen/Qwen3-VL-2B-Instruct / Qwen/Qwen3-VL-4B-Instruct	checkpoints/SPaTS-{2B,4B}-SFT
2. SPaSO	configs/spaso/spaso_2b.yaml / spaso_4b.yaml	SFT checkpoint	checkpoints/SPaTS-*-GRPO

Run SFT first, then SPaSO. Each stage reads model_name_or_path from the previous stage's output_dir.

Edit a YAML file to change training hyperparameters. Launch scripts only set distributed runtime options such as NPROC and MASTER_PORT.