Fix dense_to_sparse backward pass for three-dimensional input - #10758

Fixes #10318.

Problem

dense_to_sparse breaks autograd when the adjacency matrix is three-dimensional:

adj = torch.rand(1, 5, 5, requires_grad=True)
edge_index, edge_attr = dense_to_sparse(adj)
edge_attr.sum().backward()
# RuntimeError: one of the variables needed for gradient computation has been
# modified by an inplace operation: [torch.LongTensor [25]] is at version 2;
# expected version 0 instead.

The equivalent 2D call works fine, which is the clue.

Cause

In the 3D branch of torch_geometric/utils/sparse.py:

edge_index = flatten_adj.nonzero().t()
edge_attr = flatten_adj[edge_index[0], edge_index[1]]   # saves the index tensors
...
edge_index[1] += offset[edge_index[0]]                  # mutates them in place

The advanced indexing on the second line saves edge_index[0] and
edge_index[1] for the index backward. The in-place += then bumps their
version counter, so autograd refuses to run backward. The 2D branch never
mutates edge_index, which is exactly why only the 3D path is affected.

Fix

Build the shifted edge_index out-of-place. One line, no API change, no
behavioural change.

I checked that the forward output is unchanged rather than assuming it: for a
seeded 3 x 4 x 4 input, both edge_index and edge_attr are identical to
main on the unmasked path and on the mask= path.

Test plan

• New test_dense_to_sparse_backward in test/utils/test_sparse.py,
  parametrised over mask=None and a real mask. Both cases fail on main with
  the RuntimeError above and pass with this change.
• pytest test/utils/test_sparse.py: 20 passed.
• flake8, isort --check-only and yapf --diff are clean on both changed
  files.