Learning Neural PDE Solvers with Parameter-Guided Channel Attention