fix(python): cast uint64 row IDs to int64 before torch.from_numpy

[fightBoxing]

Summary

Fix TypeError: can't convert np.ndarray of type numpy.uint64 when using LanceDataset with with_row_id=True.

Closes #2803

Root Cause

The _rowid column from Lance scanner is pa.uint64(). In _to_tensor(), the code converts Arrow arrays to PyTorch tensors via:

tensor = torch.from_numpy(arr.to_numpy(zero_copy_only=False))
if uint64_as_int64 and tensor.dtype == torch.uint64:
    tensor = tensor.to(torch.int64)

On PyTorch versions that don't support torch.uint64 (< 2.1), torch.from_numpy() raises TypeError before reaching the cast logic.

Fix

Cast uint64 to int64 in numpy space before calling torch.from_numpy():

nparr = arr.to_numpy(zero_copy_only=False)
if uint64_as_int64 and nparr.dtype == np.uint64:
    nparr = nparr.astype(np.int64)
tensor = torch.from_numpy(nparr)

This works on all supported PyTorch versions (>= 2.0).

Testing

Added test_row_id_uint64_converts_to_int64 to verify _rowid column converts to torch.int64 when with_row_id=True.