perf(render): apply regions directly with prepared fingerprints

[TypeThe0ry]

What changed

• Replaces hot-path queued action objects and captured lambdas with five fixed-priority region passes.
• Preserves cleanup, priority order, budget boundaries, deferred retry semantics, and planned/processed/remaining metrics.
• Precomputes exact hand, discard, meld, wall, and shared-region fingerprints off the region thread.
• Adds planning-time, skipped-region, actual-update, and deferred metrics.
• Adds differential coverage for ordering and budget boundaries.

Why

Every refresh built multiple temporary lists and hundreds of captured actions, while exact FNV fingerprints were recomputed on the region thread.

Impact

Region keys, exact FNV values, entity specifications, update order, and one-tick deferred behavior remain unchanged.

Evidence

Eight paired runs:

• stable apply: 8/8 wins, about 42.6% lower time and 48.3% lower allocation
• plan/defer: 8/8 wins, about 36.6% lower time and 78.3% lower allocation

Validation

• TableRegionDisplayCoordinatorTest
• jmhJar
• spotlessCheck
• full integration test suite passed before commit splitting

[Arbousier1]

Formal trusted A/B run did not pass: INCONCLUSIVE_NOISE_OR_NO_GAIN. Evidence: https://github.com/EllanServer/MahjongEngine/actions/runs/29645451416. Closing per the performance gate policy.