Fine-grained attention in hierarchical transformers for tabular time-series