Use small test dataset (48MB) to avoid OOM

Reduce dataset size: shorter history, smaller episodes for memory efficiency

Add torch and dependencies

Add training app for CPU-basic

initial commit