write_lance with blob metadata in create mode fails

[jackye1995]

more details in lance-format/lance#5180

[chenghao-guo]

Maybe not quite related with current issue, but actually current lance-ray cannot handle large_binary blob correctly if we do not manually call the "take_blobs". A simple case can be

    blob_values = [b"foo", b"bar", b"", None, b"\x00\x01\x02"]
    ids = pa.array([0, 1, 2, 3, 4], pa.int64())

    schema = pa.schema(
        [
            pa.field(
                "blob",
                pa.large_binary(),
                metadata={"lance-encoding:blob": "true"},
            ),
            pa.field("id", pa.int64()),
        ]
    )
    table = pa.table([pa.array(blob_values, type=pa.large_binary()), ids], schema=schema)

    # Write via lance-ray
    ds_ray = ray.data.from_arrow(table)
    lr.write_lance(ds_ray, str(path), schema=schema)

Then

   # read back and write back the data to anther lance
  ds = read_lance(path)

  #write to another lance dataset in path2
  write_lance(ds, path2)

The large_binary blob will then be missing in ds2 because we didn't manually handle that.

To make it smoothly for handling blob without missing data, I am going to submit a PR that add an enhanced reader which detects Lance blob-encoded columns and reconstructs raw bytes using the LanceDataset.take_blobs API, returning LargeBinary arrays instead of the default struct<position: uint64, size: uint64>.