Use pyarrow.compute.unique on struct arrays once a struct kernel ships #170

Tracking issue for a small follow-up on PR #167.

Inside the now-removed eager duplicate-detection path we used a Python set over per-row dim tuples instead of pyarrow.compute.unique on a StructArray because pc.unique does not have a struct kernel in pyarrow 23.0.0:

pyarrow.lib.ArrowNotImplementedError: Function 'unique' has no kernel matching input types (struct<lat: double, lon: int64>)

When the dependency floor moves and pyarrow adds the kernel, the Python-side scan can be replaced with a single pc.unique(pa.StructArray.from_arrays([table.column(d) for d in dimension_columns])) and a count check.

Currently the eager reshape path no longer exists in the lazy backend (PR #167 unified eager and lazy), but the same pattern may resurface if a future feature reintroduces row-set uniqueness checks on the SQL output (e.g. for stricter duplicate diagnostics on aggregation results). Filing here so it isn't lost.