Linear speedup in saddle-point escape for decentralized non-convex optimization