Faster unique rows

[Huite]

Fixes #46

[Huite]

@veenstrajelmer

• Replaced all np.unique(..., axis=0) calls with a faster custom implementation.
• Rationale: unique relies on sorting; our implementation minimizes sorting operations.
• The order of generated connectivities (e.g., edge_node_connectivity) will differ from previous versions.
• When comparing datasets generated before and after this change, use reindex_like to ensure consistent ordering.
• Xugrid will still respect existing edge_node_connectivities, this PR doesn't change that behavior
• Expected speed improvements, among things merging the different topologies of the partitions: would be interesting to test

[Huite]

Based on these results: the sorting isn't taking enough time to make this worthwhile. (I could've known this beforehand if I had profiled properly, but it's nice to have realistic tests anyway!)