Add README with architecture details and usage (#2)

Add ram-18m training script (18,290,304 params, LLaMA-style GQA+SwiGLU) (#1)

initial commit