Release 3.0.0 RC1

Changes from 2.23.1 to 3.0.0-rc.1

• This release introduces support for variable-length chunks and variable-length
  blocks, which is the main reason for the major version bump.

  Until now, a schunk/frame generally assumed that all chunks shared the same
  logical chunk size, and regular Blosc2 chunks assumed fixed-size internal blocks
  (except for the last remainder block). In 3.0.0-rc.1, schunks can switch to
  variable chunk sizes when needed, and there is also a new chunk layout for
  variable-length blocks (VL-blocks), where each block can carry a different
  uncompressed size inside the same chunk.

  This is especially useful for workloads made of naturally variable-size pieces
  of data, like strings, records, JSON fragments, or other irregular payloads
  that previously had to be padded, split awkwardly, or stored as independent
  chunks. The new layout keeps these pieces grouped together while still making
  them individually recoverable.

  Together with this, there are new public APIs for VL-block chunks:
  blosc2_vlcompress_ctx(), blosc2_vldecompress_ctx(),
  blosc2_vlchunk_get_nblocks(), blosc2_vldecompress_block_ctx(), and
  blosc2_schunk_get_vlblock(). Lazy loading also works with VL-block chunks,
  so individual blocks can be fetched on demand without materializing the whole
  chunk first.

• The chunk and cframe formats have been extended to represent variable chunk
  sizes, VL-block chunks, and dictionary usage more explicitly. Forward
  compatibility checks were tightened as part of this work, and regular chunks
  keep their previous stable format version while VL-block chunks use a new one.

• Dictionary compression has been expanded and improved:
  use_dict now works with LZ4 and LZ4HC in addition to ZSTD, the dictionary
  state is preserved correctly across chunk compression/decompression, and the
  frame metadata now round-trips the dictionary setting. There is also a new
  minimum useful dictionary threshold to avoid training or using dictionaries
  that are too small to help.

• The necessary changes for accommodating all these improvements have been fully
  documented in README_CHUNK_FORMAT.md and README_CFRAME_FORMAT.md. Again,
  care has been taken to ensure that the chunk and frame formats are backward
  compatible with previous versions of C-Blosc2.

• Fixed several safety issues in the core library, as well as in ndcell/ndlz
  plugins. Thanks to Trail of Bits (in collaboration with Anthropic).

• Additional compatibility and portability fixes include better protection
  against unsupported future chunk/frame versions and proper VSX shuffle support
  detection on big-endian ppc64. Thanks to @AutoJanitor for these.