[evertlammerts]

Fixes the cp311 Windows wheel job of the Release workflow, which fails every run since #700 in test_a_dictionary_is_decoded_once_per_large_batch before any DuckDB code runs:

pyarrow.lib.ArrowIndexError: Index -1074 out of bounds

The test builds its dictionary indices as np.arange(2_000_000) * 1_000_000 // 2_000_000. On Windows with numpy below 2, np.arange defaults to int32 (the C long), so the product overflows and about half the indices come out negative. The cp311 job is the only Windows job that installs numpy below 2 (the tensorflow pin), which is why neither the PR workflow nor the other wheel jobs saw it. An int32 simulation gives the same minimum of -1074.

The arange now carries dtype=np.int64. Same class as #696; the other np.arange calls in the tests are small, float, or typed.