depthz: post-review cleanup pass

Comments trimmed to the load-bearing invariants (no review-history
references); no behavior change except two allocation fixes:

- zstd_append compresses into a reused thread-local buffer instead of
  resizing the outgoing message to ZSTD_compressBound, removing a
  multi-MB alloc + zero-fill per frame and right-sizing the published
  message's capacity.
- The publisher's padded-row repack buffer is thread_local scratch.

Dedup/altitude: single cpu_has_avx2() helper for all SIMD dispatch
sites; shared raster_pred_u32 for fpred encode/decode; publisher
declare-parameter boilerplate folded into a lambda; test helpers
(append_le, qpred_tol) replace repeated inline forms.

Benchmark: messages are built once and restamped per publish so the
timed loop no longer copies multi-MB payloads (better measurement
fidelity); main() split into run() (drops the gotos); dead topic
filter removed from extract_frames.py.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: facontidavide