Introduce sampling strategies

[tizianocitro]

Summary

Dataset.__getitem__ was hardcoded to sample by node as __len__ returned num_nodes, and each item returned all hyperedges incident to a given node:

• Batch count was tied to node count, so adding negative hyperedge samples had no effect on __len__.
• Uneven and unpredictable hyperedge coverage per batch due to variable node degrees.

I introduced a configurable sampling strategy that allows Dataset to sample either by node (existing behavior) or by hyperedge (new).

Hyperedge sampling (now default option):

• __len__ returns num_hyperedges.
• __getitem__(i) returns the i-th hyperedge and all its incidences, so that each batch contains a fixed, predictable number of hyperedges.
• No deduplication needed since each hyperedge appears exactly once.

Node sampling:

• __len__ returns num_nodes.
• __getitem__(i) returns all hyperedges incident to node i (and all nodes in them).
• Deduplication in Loader.collate handles any overlap.