Support reading FixedShapeTensor pyarrow columns as TensorExtensionArray

Currently pyarrow FixedShapeTensor columns are mapped to pandas ArrowExtensionArray columns (`pd.ArrowDtype`) by `read_parquet` and `from_pyarrow`, rather than to our TensorExtensionArray.

Suggested change: replace `types_mapper=pd.ArrowDtype` in `from_pyarrow` with a mapper that returns `TensorDtype` for `pa.FixedShapeTensorType` and `pd.ArrowDtype` otherwise. `TensorDtype.__from_arrow__` already handles the rest. Round-trip tests through both the pyarrow and datafusion engines, including a file whose tensor type has an identity permutation (what `pa.FixedShapeTensorArray.from_numpy_ndarray` writes; the dtype normalises it away).

Follows on from #534 and is a prerequisite for #544.