fix/reduce cuda stream syncs in bipartite_subgraph()

[cathalobrien]

Hello,

this PR reduces the number of CUDA stream syncs when running torch_geometric/utils/_subgraph.py bipartite_subgraph() on GPU from 4 to 1.
This is done by replacing the use of nonzero() with nonzero_static() and replacing an indexing operation in torch_geometric/utils/mask.py index_to_mask() with scatter()

there is still 1 more sync, i think it is this indexing here which calls nonzero()

#torch_geometric/utils/_subgraph.py bipartite_subgraph()
    edge_index = edge_index[:, edge_mask]
    edge_attr = edge_attr[edge_mask] if edge_attr is not None else None

the unit tests test/utils/test_mask.py and test/utils/test_subgraph.py pass.

See the screenshots below of pytorch perfetto traces before and after. This is accompanied by a speedup in my use case