Parallelize automatic section-search with workers

[0hmX]

Motivation

• Reduce wall-clock time of the automatic section-mask search by parallelizing candidate evaluation across worker threads.
• Avoid expensive search overhead on easy circuits by skipping the automatic search when the baseline max-region cost is already low.
• Improve final section selection by merging non-overlapping improving candidates and choosing the best merged mask.

Description

• Adds worker-based parallel search by introducing parallelSectionSearchWorker.ts and spawning Worker threads from TinyHyperGraphSectionPipelineSolver, with shared memory helpers (shareInt32Array, shareFloat64Array, shareInt8Array).
• Implements candidate chunking, worker lifecycle (startCandidateChunkWorker, finishCandidateChunkWorker), and result aggregation including ImprovementEntry and non-overlapping selection via selectNonOverlappingImprovements and getMergedMask.
• Adds a low-baseline-cost guard (minBaselineMaxRegionCostToSearch config + DEFAULT_MIN_BASELINE_MAX_REGION_COST_TO_SEARCH) that causes the search to be skipped and records sectionSearchSkipped in solver state and stats.
• Changes problem/port mask handling to use shared buffers for worker use and avoids unnecessary array copying in createProblemWithPortSectionMask for the main thread.
• Ensures pipeline output selection prefers the better of solveGraph and optimizeSection outputs and falls back to the baseline when search was skipped.
• Records additional timing and bookkeeping stats for generated/eligible/duplicate candidates and whether the search was skipped.

Testing

• Updated tests/solver/section-solver.test.ts and added section pipeline skips automatic search on low-baseline-cost circuits to assert skipping behavior and stats; the test suite covering section-pipeline behavior was executed locally and passed.
• Existing section-pipeline tests (including multi-mask search and MAX_HOT_REGIONS handling) were run and passed.

---

Codex Task

[seveibar]

this can't run in a browser