Fletcher

------------------

* Allow NumPy ufuncs to work with `np.ndarray` outputs where operations are clearly defined (i.e. the fletcher array has no nulls).

------------------

* Fix return values for `str` functions with `pandas=1.2` and `pyarrow=1`.
* Ensure that parallel variants of `apply_binary_str` actually parallize.

------------------

* Add tests for all `str` functions.
* Fix tests for `pyarrow=0.17.1` and add CI jobs for `0.17.1` and `1.0.1`.
* Implement a faster take for list arrays.
* Use `utf8_is_*` functions from Apache Arrow if available.
* Simplify `factorize` implementation to work for chunked arrays with more or less than a single chunk.
* Switch to `pandas.NA` as the user-facing null value
* Add convenience function `fletcher.algorithms.string.apply_binary_str` to apply a binary function on two string columns.

------------------

* Return correct index in functions like `fr_str.extractall`.

------------------

* Create a shallow copy on `.astype(equal dtype, copy=True)`.
* Import `pad_1d` only in older `pandas` versions, otherwise use `get_fill_func`
* Handle `fr_str.extractall` and similar functions correctly, returning a `pd.Dataframe` containing accoring `fletcher` array types.

------------------

* Use `binary_contains_exact` if available from `pyarrow` instead of our own numba-based implementation.
* Provide two more consistent accessors:
* `.fr_strx`: Call efficient string functions on `fletcher` arrays, error if not available.
* `.fr_str`: Call string functions on `fletcher` and `object`-typed arrays, convert to `object` if no `fletcher` function is available.
* Add a numba-based implementation for `strip`, `slice`, and `replace`.
* Support `LargeListArray` as a backing structure for lists.
* Implement `isnan` ufunc.