Changes from 3.3.5 to 3.4.0

This is a minor release introducing an opt-in Blosc1-compatible chunk writer,
bounded-memory APIs for variable-length block decompression, as well as
important bug fixes and robustness improvements in DELTA filter decoding, blocksize
tuning, ZFP codec handling, b2nd slicing, and thread management.

• New opt-in writer for Blosc1 (C-Blosc 1.x compatible) chunks:
  blosc2_ctx_set_blosc1_format() makes a compression context write chunks in
  format version 2 with the 16-byte header, which C-Blosc 1.x can decode (stock
  C-Blosc2 chunks, even with BLOSC_BLOSC1_COMPAT, cannot be). Besides the
  header, such chunks follow the version 2 bitshuffle rule (a block is
  bit-shuffled only if its number of elements is a multiple of 8) and only split
  blocks that a C-Blosc 1.x reader also splits. Only plain shuffle/bitshuffle with
  the blosclz, lz4, lz4hc, zlib and zstd codecs are supported; anything else
  (other filters including delta, other codecs, dictionaries, variable-length
  blocks) fails with an error. Off by default, no ABI change. The
  BLOSC_BLOSC1_COMPAT environment variable now enables the same mode for
  blosc1_compress() (it previously only dropped the extended header, still
  writing a format version 5 chunk), so delta there is now an error.
  blosc2_ctx_get_blosc1_format() reports whether a context is in this mode. The
  filter may be in any slot of the pipeline; a prefilter is applied as usual.
  Thanks to Mark Kittisopikul (@mkitti).
• Added blosc2_vlblock_get_nbytes_ctx() to query one block's uncompressed size
  without payload, dictionary, or scratch allocation. Lazy queries read only
  the selected four-byte prefix.
• Added blosc2_vldecompress_block_into_ctx() for single-block decompression
  into a caller-owned buffer. Capacity is checked before dictionary/scratch
  allocation, and scratch is sized from the admitted block. Both APIs support
  ordinary and lazy VL chunks. The existing allocating API is unchanged.
• Fixed DELTA filter decoding on partial chunks, maskout, and getitem:
  When an array compressed with the DELTA filter was sliced or read partially,
  non-zero blocks could be corrupted because DELTA encodes blocks > 0 relative
  to block 0. If block 0 was masked out or omitted (such as in b2nd's compact
  read path for small block fractions, or in block-level getitem), subsequent
  blocks decoded against uninitialized memory. b2nd now disables the compact
  path for DELTA and always unmasks block 0, and _blosc_getitem and
  blosc2_decompress_block_ctx fall back to full chunk decompression when DELTA
  is active. Destination pointers and sizes are reset after fallback decompression
  to avoid leaving dangling references on reused contexts. Fixes Blosc/python-blosc2#740.
  Thanks to Karol Gotkowski (@Karol-G) for reporting this issue.
• Fixed bit-shuffled Blosc1 chunks with leftover bytes being decoded incorrectly:
  When reading a chunk with format version 2 whose bit-shuffled blocks have a
  number of elements that is a multiple of 8, the bytes left over after the last
  whole element were not copied to the output, so the end of the buffer was
  garbage while the call reported success. They are now copied, as C-Blosc 1.x
  does. See #812. Thanks to Mark Kittisopikul (@mkitti).
• Fixed chunks that could not be read back when the requested block size was not a
  multiple of the typesize: the tuner rounds a requested block size down to a multiple
  of the typesize, but the chunk header recorded the size that had been asked for, so a
  reader cut the chunk into blocks of a different size and failed (error -3, or wrong
  data) with every filter and codec. The header now records the block size that is used.
  Thanks to Mark Kittisopikul (@mkitti).
• Fixed the block size of buffers with a typesize above BLOSC_MAX_TYPESIZE: such
  buffers are stored as a stream of bytes with a typesize of 1, but the tuner chose the
  block size with the large typesize and only then was it capped. A buffer shorter than
  the typesize was split into blocks of one byte (200 bytes of typesize 300 took 1831
  bytes) and a longer one got a block size that was a multiple of the large typesize.
  The typesize is now capped before the tuner runs, so that it is blocked exactly like a
  typesize of 1. Thanks to Mark Kittisopikul (@mkitti).
• blosc2_compress_ctx() with a typesize of 0 stopped with SIGFPE (a division by
  zero when choosing the block size). It now returns BLOSC2_ERROR_INVALID_PARAM.
  Thanks to Mark Kittisopikul (@mkitti).
• Hardened ZFP codec against bitstream over-read and invalid metadata:
  ZFP bitstream decompression now operates from a zero-padded buffer sized to
  zfp_stream_maximum_size(), preventing out-of-bounds heap reads when 64-bit
  word reads extend past unaligned payloads (blocks * bits not a multiple of 64).
  Also guarded against word-alignment over-read in zfp_getcell(), removed
  redundant stream opens, validated metalayer bounds via b2nd_deserialize_meta_inline,
  and rejected non-positive blockshapes and invalid ndim.
  Thanks to @kalt2212 and Farkhalit Rida (@fr-manvi).
• Hardened b2nd_get_slice_nchunks bounds handling and allocation:
  Clamped slice bounds against negative values and extshape, sized chunks_idx
  from slice bounding boxes to prevent mismatch overflows, protected against
  multiplication overflow on 32-bit platforms, and handled empty slices and
  shrinking reallocations without memory leaks.
  Thanks to Farkhalit Rida (@fr-manvi).
• Rejected split blocks where streams do not cover blocksize:
  In blosc_d, decompression verifies that neblock * nstreams == bsize and returns
  BLOSC2_ERROR_DATA if split streams fail to cover the full block, avoiding leaving
  bytes unwritten in destination buffers (backported from C-Blosc #407).
  Thanks to @suraiya-ctrl.
• Avoided unused parallel work for small block counts:
  Single-block compression and decompression jobs remain serial when nthreads > 1,
  avoiding worker thread pool startup and overhead when threads cannot be used.
  Parallel dispatch also caps worker entries to min(nthreads, nblocks).
  Fixed thread backend initialization on Windows. Thanks to
  Johnny Kao (@Johnny-Kao).
• Added examples/vlchunks.c illustrating variable-length chunk usage, updates,
  compression, and verification. Thanks to Lance Uymatiao (@lanceu0128).