Implicit Temporal Modeling with Learnable Alignment for Video Recognition