Fix preparing regions for numpy ndarray.

[shaunrd0]

This fixes hanging task graph nodes in TileDB-Server, due to regions being cleared when passed as a numpy.ndarray.

After clearing the regions, we would later select all regions for the read within prepare_regions_v4, resulting in each task graph node reading the entire dataset and eventually hitting a timeout after 15m.

The reproduction in CLOUD-2882 completes normally with max_workers=10 when running locally against this branch.

[alancleary]

This PR currently only updates 1 of 4 places a regions parameter is parsed in this way. All 4 should be updated:

1. def read_arrow()
2. def read()
3. def export()
4. def read_iter()

[alancleary]

Switching to elif prevents _prepare_regions(...) from being called in the case that regions is a string, meaning the region is no longer parsed and validated. Apparently there isn't a test for this so I ask that you add this test to the PR and revise your approach, for example:

with pytest.raises(Exception, match=format_error):
    test_stats_v3_ingestion.read_arrow(regions="", ...)

[alancleary]

The else should be left as the base case so that types that aren't explicitly supported don't make it past this block. In other words, add a new elif for type numpy.ndarray and apply the _prepare_regions(...) generator or something similar if required to preserve the numpy.ndarray type.