VaLID: Variable-Length Input Diffusion for Novel View Synthesis