Average Gradient Outer Product in kernel regression provably recovers the central subspace for multi-index models