Fix RuntimeError in maybe_num_nodes_dict with empty edge indices from #10386

[DhyeyMavani2003]

Description

This PR fixes a bug where maybe_num_nodes_dict would crash with a RuntimeError when processing heterogeneous graphs containing empty edge indices.

Problem

The function calls .max() on edge index tensors without checking if they are empty. PyTorch requires specifying a reduction dimension when calling .max() on empty tensors, causing the following error:

RuntimeError: max(): Expected reduction dim to be specified for input.numel() == 0. Specify the reduction dim with the 'dim' argument.

Reproducer

import torch
from torch_geometric.utils.num_nodes import maybe_num_nodes_dict

edge_index_dict = {
    ('user', 'rates', 'movie'): torch.tensor([[], []], dtype=torch.long),
    ('user', 'follows', 'user'): torch.tensor([[0, 1], [1, 2]], dtype=torch.long),
}

# This would crash before the fix
result = maybe_num_nodes_dict(edge_index_dict)

Solution

Added checks for empty tensors (numel() > 0) before calling .max(), returning 0 for empty edge indices, which is the correct behavior for graphs with no edges of a particular type.

Impact

This bug affects users working with heterogeneous graphs where some edge types have no edges - a common scenario in real-world applications:

• Recommendation systems (users who haven't rated certain items)
• Knowledge graphs (missing relationship types)
• Social networks (disconnected components)

Testing

Added comprehensive test coverage in test/utils/test_num_nodes.py:

• Mixed empty and non-empty edge indices
• All empty edge indices
• Empty edge indices with provided num_nodes_dict

All tests pass successfully.

Checklist

• Bug fix identified and reproduced
• Fix implemented with proper edge case handling
• Comprehensive tests added
• All tests pass
• Commit message follows repository conventions