Pseudorandom Streams within Diffusion Models Act as Learnable Inputs That Affect Generation Quality