0.3.0

Major changes:

• We now provide two different extension array implementations.
  There now is the more simpler FletcherContinuousArray which is backed by a pyarrow.Array instance and thus is always a continuous memory segments.
  The initial FlectherArray which is backed by a pyarrow.ChunkedArray is now renamed to FletcherChunkedArray.
  While pyarrow.ChunkedArray allows for more flexibility on how the data is stored, the implementation of algorithms is more complex for it.
  As this hinders contributions and also the adoption in downstream libraries, we now provide both implementations with an equal level of support.
  We don't provide the more general named class FlectherArray anymore as there is not a clear opinion on whether this should point to FletcherContinuousArray or FletcherChunkedArray.
  As usage increases, we might provide such an alias class in future again.
• Support for ArithmeticOps and ComparisonOps on numerical data as well as numeric reductions such as sum.
  This should allow the use of nullable int and float type for many use cases.
  Performance of nullable integeter columns is on the same level as in pandas.IntegerArray as we have similar implementations of the masked arithmetic.
  In future versions, we plan to delegate the workload into the C++ code of pyarrow and expect significant performance improvements though the usage of bitmasks over bytemasks.
• any and all are now efficiently implemented on boolean arrays.
  We blogged about this and how its performance is about twice as fast while only using 1/16 - 1/32 of RAM as the reference boolean array with missing in pandas.
  This is due to the fact that prior to pandas=1.0 you have had to use a float array to have a boolean array that can deal with missing values.
  In pandas=1.0 a new BooleanArray class was added that improves this stituation but also change a bit of the logic.
  We will adapt to this class in the next release and also publish new benchmarks.

New features / performance improvements:

• For FletcherContinuousArray in general and all FletcherChunkedArray instances with a single chunk, we now provide an efficient implementation of take.
• Support for Python 3.8 and Pandas 1.0
• We now check typing in CI using mypy and have annotated the code with type hints.
  We only plan to mark the packages as py.typed when pandas is also marked as py.typed.
• You can query fletcher for its version via fletcher.__version__
• Implemented .str.cat as .fr_text.cat for arrays with pa.string() dtype.
• unique is now supported on all array types where pyarrow provides a unique implementation.