Small bundled fixes: #153, #94, #134, partial #64

[nclack]

Small, contained fixes for several open issues. One commit per issue so the
PR stays reviewable.

Commits

• Guard flush against unsized layout — Closes Divide-by-zero in stream_flush_body when destroying a half-created GPU stream #153.
  stream_flush_body divided by ctx->layout.epoch_elements, which is 0 when
  tile_stream_gpu_create fails before binding the array (an engine_limits
  or stream_engine_init failure, both of which run before
  engine_array_state_init sets ctx->layout at src/gpu/stream.init.c:221).
  The destroy auto-flush then hit a SIGFPE. The guard early-returns when
  nothing was ever sized. (Justified by code-walk in the commit; a clean unit
  test would need to force a GPU init failure deterministically, which is
  GPU-dependent and fragile.)

• Test uniform chunk bytes across LODs — references benchmark compression ratios for multiscale are not accurate #94.
  Adds a CPU-only test (test_uniform_chunk_bytes_across_levels in
  tests/test_lod_cpu.c) that builds a multiscale layout via
  compute_stream_layouts and asserts every level's
  agg_layout.max_comp_chunk_bytes is identical and equals the L0 chunk byte
  size (chunk_stride * bpe) for CODEC_NONE. Locks in the invariant the
  runtime relies on (batch_aggregate_layout_init asserts it;
  compute_stream_layouts feeds one max_output_size to every level).

• Drop fill_value from ngff config — Closes fill_value is a zarr thing not an ngff thing #134.
  fill_value is a per-array zarr property: it is written to zarr array
  metadata (src/zarr/zarr_metadata.c:201) and already lives on
  zarr_array_config. The OME-NGFF multiscale config exposed a redundant
  copy. Removed it from ngff_multiscale_config; ngff-created arrays now take
  the zarr default. Every caller passed 0, so behavior is unchanged. The zarr
  metadata test (test_json_writer.c) asserts fill_value in zarr metadata
  (correct placement) and needed no change.

• Emit memory estimate in bench JSON — references benchmark sweep: record memory usage #64 (estimate half only).
  The per-stream memory estimate is already computed and printed by the bench;
  this emits it into the JSON as two additive fields,
  memory_estimate_total_bytes (device bytes on GPU, heap bytes on CPU) and
  memory_estimate_pinned_bytes (pinned host bytes on GPU, 0 on CPU), so the
  sweep records it. Additive only: scripts/sweep/models.py
  (RunResult(extra="allow")) validates both old result files and freshly
  emitted ones — verified against both. The "actual RSS vs baseline" half of
  benchmark sweep: record memory usage #64 is out of scope, so benchmark sweep: record memory usage #64 stays open.

Dropped from the bundle

• Per-LOD timing metrics can race the next batch's events (skew only, no data impact) #154 (per-LOD timing-metric skew) — DROPPED, not contained.
  The issue assumes a fix that snapshots/double-buffers the per-fc LOD timing
  "the same way the aggregate timing is already handled." On inspection that
  pattern does not transfer: the aggregate timing events are recorded at kick
  time and are safe only because the drain-before-rekick host rule keeps the
  same fc's slot drained before it is re-kicked. The LOD timing events are
  recorded at epoch-fill time (lod_run_epoch), which is not gated against
  the prior same-fc drain, so the identical handle snapshot would be a no-op
  against the value race (the snapshot copies a handle to the same event object
  the producer then re-records). Worse, t_end is dual-purpose — it is also
  the GPU_EDGE_LOD_DONE synchronization edge (bound at
  src/gpu/stream.init.c:177) — so relocating or re-keying these events
  touches load-bearing ordering, not just metrics. A correct fix needs extra
  timing generations or a new drain-before-refill gate, i.e. a new mechanism
  rather than a handful of lines. Since the bug is cosmetic (metrics-quality
  only, never wrong output bytes), that cost is not worth risking this PR.
  Left open with this rationale.

Validation

Built RelWithDebInfo with GPU on (sm_89, L40); ctest -E "(s3)" green; no new
build warnings; the new invariant test passes; the bench JSON emits the new
fields on both GPU and CPU runs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/ngff/ngff_multiscale.c	58.79% <ø> (-0.18%)	⬇️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.