LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias