Support multiple features

[zheng-da]

Currently, when a graph has multiple features, GraphStorm concatenates all features into a single tensor before feeding them to the model for training. This assumes that all features have the same data types. This is problematic if different node data have different data types.