Changes from 3.2.1 to 3.2.2

Performance improvements

• Avoid chunk-sized scratch allocation for small b2nd get_slice
  requests. get_set_slice() in the b2nd path previously malloc'd a
  full extended-chunk-sized scratch buffer on every call, even when the
  block maskout meant only one block was decompressed into it. For large
  chunks the mmap/munmap page-table work alone is O(chunksize), so reading
  a single 1.6 MB block from a 128 MB chunk could cost up to 3x more than
  the same read from a small chunk. Now the get path counts the needed
  blocks after building the maskout, and when they are a small part of the
  chunk, decompresses just those blocks one at a time via
  blosc2_getitem_ctx() into a reusable block-sized buffer — small
  reads stay O(request) instead of O(chunksize), and on-disk frames only
  read the touched blocks. Larger requests keep the existing parallel
  block-decompression path unchanged, and the chunk-sized scratch is now
  allocated lazily only when that path (or a set operation) actually runs.
  Single-block reads from large chunks are now flat ~0.33 ms regardless
  of chunk size.

Documentation

• Broad refresh: annotated every ROADMAP-TO-3.0 item as Done/Ongoing/
  Deferred now that 3.0 is out, and added a new ROADMAP-TO-4.0 with the
  deferred items as current goals. Fixed several stale references
  (python-blosc2 status, NEON bitshuffle, citation year, Twitter→Mastodon,
  master→main, header paths), and assorted grammar/markup fixes.

Notes

• This is a maintenance release with no API/ABI changes.