Changes from 3.9.1 to 3.9.2

Bugfixes

Fix the assembly of returned slice data in Blosc2 NDim optimized slice reads by using Blosc2's b2nd_copy_buffer (#1078). The bug only showed up when the chunk did not fully cover the innermost dimension. Add unit tests to ckeck for regressions, along with foreign-generated files, and enable and fix Blosc2 NDim tests which were not being run. Thanks to Ivan Vilata.

PyTables wheels now use a threadsafe build of the HDF5 library (#1075 and #1077). Concurrent reads should be possible with no need for additional locking or monkey-patching of the file open function. Thanks to Kiet Pham.
Partial support for the future NumPy 2, with some tests still failing (#1068). Thanks to Thomas Grainger.
Relax the reading of Blosc2 NDim to cope with datasets stored with other tools (#1072), e.g. missing chunk rank/shape in filter values, having HDF5 chunks where the Blosc2 super-chunk contains more than one inner chunk, or chunks with data not padded to the full chunk size (example script and tests included). Also enhance checks, comments and logged messages. Thanks to Ivan Vilata.

Drop compatibility with the obsolete HDF5 1.8 API. PyTables now requires at least the 1.10 API (#1080). Thanks to Antonio Valentino.
Require python-blosc2 >= 2.3.0 or c-blosc2 >= 2.11.0 (which adds support for the b2nd_copy_buffer function).
Use the main Conda Forge channel for Python 3.12 (#1066). Thanks to Thomas Grainger.
Assorted fixes to the b2nd slicing benchmark. Thanks to Ivan Vilata.
Assorted fixes to b2nd slicing optimization tips (#1069). Thanks to Ivan Vilata.

In alphabetical order: