On Principles, Models and Methods for Learning from Irregularly Sampled Time Series: From Discretization to Attention and Invariance