fix(python): support batched UInt8 Hamming queries

[ddupg]

Summary

Batched NumPy uint8 queries were converted to List<Float32> and skipped the existing UInt8 cast, causing Hamming search to fail with a type mismatch.

Preserve the batch shape while casting its elements back to UInt8.

Reproduction

import lance
import numpy as np
import pyarrow as pa
import tempfile

with tempfile.TemporaryDirectory() as root:
    ds = lance.write_dataset(
        pa.table({
            "vector": pa.array(
                [[0, 0], [255, 0]],
                type=pa.list_(pa.uint8(), 2),
            )
        }),
        f"{root}/data.lance",
    )
    ds.to_table(nearest={
        "column": "vector",
        "q": np.array([[0, 0]], dtype=np.uint8),
        "metric": "hamming",
        "k": 1,
    })

Before this change:

Column vector has element type UInt8 and the query vector is Float32

[lance-gatekeeper]

✅ Gate recommendation: approve.

The shape-preserving UInt8 cast restores the documented batched-query contract for binary vectors while leaving scalar-query and core search semantics unchanged. Flat and IVF_FLAT batch results match repeated single-query searches.