perf: O(N) array-slice scan + primitive cursor cache

[JohannesLichtenberger]

Summary

Rebased onto current main after the temporal-axis-prefetch and OBJECT_NAMED_* leaf changes landed. The branch is now a single focused commit.

JsonDBArraySlice.getValues() was O(N²): every at(fromIndex+i) rebuilt a ChildAxis and walked from index 0. Replaced with a single forward pass that skips fromIndex siblings and then collects the slice. ArrayList pre-sized to slice width.

Sequential at(i) access was also O(N²) for the same reason. Added a primitive cursor cache (int cursorSliceIndex, long cursorNodeKey); the next monotone access advances one moveToRightSibling instead of restarting. Validated against the slice's array via getParentKey() == arrayKey so structural mutations can't surface a stale element.

Fixed an indexing bug on the cached path of at(IntNumeric) — values.get(fromIndex + i) was indexing a slice-local list with a global index, throwing IndexOutOfBoundsException whenever fromIndex > 0 after values() had been called.

Dropped the duplicate JsonItemFactory instantiation by exposing the parent's via a protected getter.

On the dropped parallel materialisation

An earlier revision of this branch added a ForkJoinPool-driven parallel path for atomic-only slices ≥ 262 144 elements, with full Stage-1/2/3 fan-out. After rebasing onto main (which brought in temporal-axis prefetching and fused OBJECT_NAMED_* leaves), the serial path's per-element cost dropped from ~200 ns to ~93 ns. At that work density, rtx-creation + ForkJoinPool fixed costs eat the entire parallel budget — measured 1.00× ratio (no payoff at all) on a 19-way commonPool. Dropped per the standing rule that parallelism in this codebase must demonstrably pay off; if it ever stops paying, remove it.

If a future change makes per-element decode expensive again (e.g., a new compression codec), revisit array-slice-parallelism.md in the project memory — the design proof and obligations are documented there.

Test plan

• ./gradlew :sirix-query:test --tests "io.sirix.query.json.JsonDBArraySliceTest" — 11 tests, 0 failures.
• Cursor cache stress (130 sequential reads on a 200-element array, then a backward jump).
• Indexing-bug regression on cached path.
• Sub-slicing semantics pinned down (existing absolute-index behaviour preserved as rangeOnSlice_yieldsSubSlice_withCorrectAbsoluteIndices).
• Bounds-throw paths covered.