Deformed semicircle law and concentration of nonlinear random matrices for ultra-wide neural networks