Support hyperedge-based sampling in `Dataset`

[tizianocitro]

Dataset.__getitem__ currently samples by node: given a node index, it returns all hyperedges incident to that node. This works for node-centric tasks but is suboptimal for HLP, where the prediction unit is a hyperedge.

Problems with node-based sampling for HLP:

• __len__ returns num_nodes, so batch count is independent of the number of hyperedges (adding negative samples doesn't change batch count)
• A sampled node pulls in all its hyperedges, giving unpredictable and uneven hyperedge coverage per batch.
• Multiple sampled nodes can pull in the same hyperedge, requiring deduplication in collate

I want to add a configurable sampling strategy to Dataset so that __getitem__ can sample either by node or by hyperedge.

When sampling by hyperedge:

• __len__ returns num_hyperedges.
• __getitem__(i) returns the i-th hyperedge (with all its incidences).
• Each batch contains a fixed number of hyperedges, giving direct control over positive/negative ratio per batch.
• No deduplication needed in collate since each hyperedge is returned exactly once.

When sampling by node (current behavior):

• __len__ returns num_nodes.
• __getitem__(i) returns all hyperedges incident to node i.
• Deduplication in collate handles overlapping hyperedges.

[tizianocitro]

Assigned this to me.