PERF: AlignSections Filters OoC Optimization

[joeykleingers]

Summary

Adds slice-buffered OOC algorithm paths for the 4 AlignSections filters, using dual-dispatch (Strategy C) to preserve the original in-core code untouched while adding OOC-optimized variants.

Changes

Base class (AlignSections.cpp)

• New AlignSectionsTransferDataOocImpl<T> dispatched when any cell array is OOC
• Reads each Z-slice into a local buffer, applies the 2D X/Y shift in memory, writes back — eliminates per-tuple chunk thrashing

AlignSectionsMisorientation

• New findShiftsOoc() buffers 2 adjacent Z-slices of quats, cellPhases, and mask before the convergence loop
• Pre-loads first reference slice, then std::swap cur→ref each iteration to avoid re-reading the reference from DataStore
• Eliminates repeated ZarrStore reads during the 7×7 candidate shift grid convergence

AlignSectionsMutualInformation

• New formFeaturesSectionsOoc() buffers one Z-slice of quats, cellPhases, and mask for the per-slice 2D flood-fill segmentation

AlignSectionsFeatureCentroid & AlignSectionsListFilter

• Benefit from transfer phase optimization only (findShifts access patterns are already sequential/trivial)

Tests

• All 9 correctness tests now exercise both in-core and OOC algorithm paths via GENERATE(false, true) + ForceOocAlgorithmGuard
• Benchmark tests (200³ programmatic datasets) left without GENERATE for clean timing

Benchmark Results (200×200×200)

Filter	In-Core Before → After	OOC Before → After	OOC Speedup
AlignSectionsMisorientation	0.74s → 0.79s (~1.0x)	32.89s → 16.14s	2.0x
AlignSectionsMutualInformation	0.49s → 0.52s (~1.0x)	15.61s → 15.14s	~1.0x
AlignSectionsFeatureCentroid	0.24s → 0.28s (~1.0x)	8.41s → 8.82s	~1.0x
AlignSectionsListFilter	0.22s → 0.25s (~1.0x)	7.50s → 7.95s	~1.0x

Optimization Ceiling Analysis

The OOC speedups are more modest than Groups B/C/D because the transfer phase dominates OOC runtime and is bottlenecked by ZarrStore's per-element overhead (~55–75ns per operator[]: mutex lock/unlock + chunk lookup vs ~1ns for in-core DataStore). The Misorientation filter shows the most benefit (2.0x) because its findShifts convergence loop re-reads the same 2 slices many times — slice buffering plus reference-slice swap (reusing the previous iteration's current-slice buffer as the next iteration's reference) eliminates that repeated I/O.

Further improvement requires deeper OOC infrastructure changes:

1. Bulk read/write API on AbstractDataStore — eliminates ~47.8M mutex lock/unlock cycles per filter (~1–2s savings)
2. Chunk-level bulk transfer in FileCore — bypasses per-element chunk lookup entirely (estimated 3–5x transfer phase improvement)
3. Larger FIFO cache or per-array cache isolation — enables parallel OOC array processing

Test Plan

• All 9 correctness tests pass on in-core build (simplnx-Rel)
• All 9 correctness tests pass on OOC build (simplnx-ooc-Rel)
• GENERATE(false, true) exercises both algorithm paths in both builds
• Benchmark tests confirm zero in-core regression