Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation