Fix unbatch for COO batches with zero-edge graphs

[Parvm1102]

closes #650

The previous implementation inferred edge boundaries by looking for the first edge of the next graph, which breaks when an intermediate graph has no edges. This could assign edges to the wrong graph during unbatch.

FIX

• replace boundary inference in _unbatch_edgemasks with direct edge bucketing by graph
• it creates one edge bucket per graph
• it iterates over edges once
• for each edge, it computes the owning graph from the source node’s interval in cumnum_nodes
• it pushes that edge into the correct bucket

TESTS

• add regression tests for zero-edge graphs in COO batches, including feature roundtrips