Changes from 3.3.3 to 3.3.4

This release introduces bounded-memory tail compaction for contiguous on-disk
frames and clean file eviction for payload-free chunks in sparse frames.

• Bounded-buffer tail compaction in contiguous frames:
  When updating a chunk in place within an on-disk contiguous frame (cframe),
  a change in compressed size requires shifting all subsequent chunks in the
  payload tail. Previously, Blosc2 allocated a single heap buffer sized to the
  entire remaining tail (malloc(tail_nbytes)). For multi-gigabyte or terabyte
  frames, updating an early chunk could require gigabytes of RAM, risking
  out-of-memory aborts.

  Tail compaction now uses an internal bounded-buffer movement helper
  (frame_move_range) capped at 1 MiB (FRAME_TAIL_COPY_BUFFER_CAP), decoupling
  RAM consumption from frame size:

  • Shifts forward when shrinking (dst < src) and backward when expanding
    (dst > src) to guarantee overlap safety across any displacement.
  • Reads each segment fully into owned scratch memory before writing, avoiding
    pointer invalidation when memory-mapped files are remapped during expansion.
  • Adds checked arithmetic to prevent integer overflow and negative buffer
    allocations.
  • Consolidates error cleanup paths in frame_update_chunk() and
    frame_update_trailer(), and accounts for frame->file_offset on truncation.

• Sparse frame special chunk file removal:
  Updating a chunk to a payload-free special value (such as ZERO, UNINIT, or NAN)
  in a sparse frame (sframe) previously truncated the chunk file to 0 bytes
  rather than deleting it. Repeated cache eviction and refill cycles could leave
  thousands of empty chunk files (0000000X.chunk), cluttering directories and
  exhausting filesystem inodes.

  Updating to a payload-free special value now unlinks and deletes the old
  physical chunk file (sframe_delete_chunk()), matching the representation of
  never-materialized special chunks:

  • BLOSC2_SPECIAL_VALUE chunks (which store repeated payload data) retain
    their physical chunk file.
  • Adopts recoverable-orphan ordering: the chunk file is deleted only after
    publishing the updated offsets index and trailer to disk.
  • Properly tracks the preceding physical chunk ID when logical chunks have been
    reordered.
  • Custom I/O backends that map file paths are handled cleanly: standard backends
    use remove() (treating ENOENT as idempotent success), while custom backends
    use their registered callbacks (open, truncate to 0, close), preventing
    unintended file operations on host paths.
  • Eviction failures occurring after the metadata update has been committed are
    logged as warnings rather than returning an error, preventing committed
    updates from being reported as failures.

• Memory mapping (mmap) fixes and external file growth preservation:

  • Memory remapping on non-Linux POSIX systems (macOS, BSD) in
    blosc2_stdio_mmap_write() now maps cleanly to a new address before
    unmapping the old one, avoiding MAP_FIXED pitfalls.
  • In blosc2_stdio_mmap_destroy(), file size truncation respects external
    growth. On Windows, file padding to mapping_size is only trimmed if the
    file has not grown externally, avoiding truncating appends made by other
    handles. On POSIX, avoids truncating away external writes or appends.

• New benchmarks and tests:

  • Added bench/bench_compact_tail.c to benchmark tail updates across
    different buffer caps and chunk positions (with portable timing support
    across POSIX and Windows).
  • Added comprehensive test suites in tests/test_compact_tail_update.c and
    tests/test_sframe_special_eviction.c.

There are no API or format changes in this release.