Changes from 3.1.5 to 3.2.0

New features

• Opt-in cross-process locking for disk-based frames, via a sidecar
  lock file (.b2lock, flock/LockFileEx) plus a generation
  counter that detects mutations other handles made even when the on-disk
  length is unchanged. Enable it per-handle (blosc2_stdio_params.locking,
  or the equivalent for other I/O backends) or fleet-wide with the
  BLOSC_LOCKING environment variable. Advisory: every handle touching
  the container must opt in. Not supported over NFS.

• blosc2_schunk_lock() / blosc2_schunk_unlock(): a bracket API to
  hold the exclusive frame lock across several operations, making a
  multi-step mutation atomic to other locked handles instead of only each
  individual call. No-ops on a handle without locking enabled.
  b2nd_resize() and b2nd_set_slice_cbuffer() now bracket their own
  multi-chunk sequences this way too, so a locked reader never observes a
  half-applied resize or slice write, and two overlapping slice writers no
  longer interleave at chunk granularity.

• Growth-SWMR (single writer, multiple readers): a reader handle
  opened on a disk-based frame or b2nd array now follows shape/length
  growth made through another handle (typically another process) on its
  next access, without reopening. New public b2nd_refresh() reader-side
  API, and examples/b2nd/example_growth_swmr.c demonstrating the
  contract. Works with or without locking; only the on-disk length is
  polled without it, so growth is virtually always noticed but a
  same-length mutation may not be until the next one that changes it.

• Stale-handle coherence fixes: a handle whose cached view of an
  on-disk frame goes stale (another handle rewrote it) now re-syncs before
  trusting cached state, closing gaps in blosc2_vlmeta_exists() and in
  the counters (nbytes/cbytes/nchunks) that
  blosc2_schunk_append_chunk()/blosc2_schunk_insert_chunk() apply their
  deltas to. Also fixed:

  • A race between opening a frame and a concurrent writer growing or
    otherwise mutating it (blosc2_schunk_open_offset_udio() could
    return NULL on a transient "frame length exceeds file boundary"
    read instead of retrying it).
  • A torn read during the (unlocked) staleness poll used to hard-error
    instead of keeping the cached view and retrying on the next poll, as
    intended.
  • A leak, and a free of an uninitialized pointer in
    blosc2_schunk_copy(), on an unreadable-frame staleness poll (hit
    legitimately by a freshly created, not-yet-flushed file-backed
    schunk).

API/ABI notes

• Appends change_tick to blosc2_schunk struct and last_tick to
  b2nd_array_t, and adds locking to blosc2_stdio_params and the new
  BLOSC2_ERROR_LOCK error code. No existing field or function signature
  was altered, but this breaks the ABI as two public structs have an
  additional field; the shared library's SOVERSION is bumped accordingly
  (8 to 9).

Notes

• Advisory locking is not a replacement for a real lock manager: it only
  protects a container if every handle on it (across every process)
  enables it. It does not work on network filesystems (NFS), and is
  incompatible with memory-mapped I/O.
• See plans/high-level-formats-locking.md and
  plans/todo-locking-swmr.md in the source tree for the full design
  rationale and known follow-ups.