What's new in h5py 3.9

This version of h5py requires Python 3.8 or above.

New features

New out argument to :meth:`~h5py.h5d.DatasetID.read_direct_chunk` to allow passing the output buffer (:pr:`2232`).
The objects from :meth:`.Dataset.asstr` and :meth:`.Dataset.astype` now implement the __array__() method (:pr:`2269`). This speeds up access for functions that support it, such as np.asarray().
Validate key types when creating groups and attributes, giving better error messages when invalid types are used (:pr:`2266`).

Using :meth:`.Dataset.astype` as a context manager has been removed, after being deprecated in h5py 3.6. Read data by slicing the returned object instead: dset.astype('f4')[:].

H5Pget_elink_acc_flags & H5Pset_elink_acc_flags as :meth:`h5py.h5p.PropLAID.get_elink_acc_flags` & :meth:`h5py.h5p.PropLAID.set_elink_acc_flags`: access the external link file access traversal flags in a link access property list (:pr:`2244`).
H5Zregister as :func:`h5py.h5z.register_filter`: register an HDF5 filter (:pr:`2229`).

Group.__contains__ and Group.get now use the default link access property list systematically (:pr:`2244`).
Removed various calls to the deprecated numpy.product function (:pr:`2242` & :pr:`2273`).
Fix the IPython tab-completion integration in IPython 8.12 (:pr:2256`).
Replacing attributes with :meth:`.AttributeManager.create` now deletes the old attributes before creating the new one, rather than using a temporary name and renaming the new attribute (:pr:`2274`). This should avoid some confusing bugs affecting attributes. However, failures creating an attribute are less likely to leave an existing attribute of the same name in place. To change an attribute value without changing its shape or dtype, use :meth:`~.AttributeManager.modify` instead.

When building with :ref:`parallel` support, the version of mpi4py used on various Python versions is increased to 3.1.1, fixing building with a newer setuptools (:pr:`2225`).
Some fixes towards compatibility with the upcoming Cython 3 (:pr:`2247`).