Default iOS rendering to Ganesh and add benchmarks

[markm39]

Summary

• Adds the iOS Ganesh/Metal render path, defaults public iOS rendering to Ganesh, and keeps CPU selectable for A/B comparisons and fallback.
• Adds on-device benchmark tooling for replayed strokes, manual notebook recordings, viewport scroll sampling, multi-page suite runs, eraser, selection, and tool coverage.
• Applies Ganesh output color compatibility so tool colors and selection chrome match the CPU path.

Benchmark evidence

Captured on iPad using this branch's example benchmark screen. Lower p95 values are better; higher scroll FPS is better.

Manual full-notebook recording

xychart-beta
  title "Manual full-notebook p95"
  x-axis ["CPU render", "Ganesh render", "CPU latency", "Ganesh latency"]
  y-axis "milliseconds" 0 --> 16
  bar [13.93, 3.76, 8.41, 6.43]

Loading

Metric	CPU	Ganesh	Result
Render p95	13.93 ms	3.76 ms	73% lower
Input-to-present latency p95	8.41 ms	6.43 ms	24% lower
Memory delta	101.9 MB	83.7 MB	18.2 MB lower
Viewport FPS	56.92	53.25	same range while scrolling/drawing

12-page automated notebook suite

xychart-beta
  title "12-page suite p95"
  x-axis ["CPU render", "Ganesh render", "CPU latency", "Ganesh latency"]
  y-axis "milliseconds" 0 --> 18
  bar [7.35, 4.87, 13.15, 16.92]

Loading

xychart-beta
  title "12-page suite scroll FPS"
  x-axis ["CPU", "Ganesh"]
  y-axis "fps" 0 --> 60
  bar [51.54, 52.68]

Loading

Metric	CPU	Ganesh	Result
Render p95 average	7.35 ms	4.87 ms	34% lower
Synthetic input latency p95 average	13.15 ms	16.92 ms	replay latency higher in the suite
Native dropped frames	134	183	higher in synthetic suite
Scroll FPS	51.54	52.68	same range
Scroll dropped frames	38	41	same range

Test plan

• npm run validate
• xcodebuild -quiet -workspace example/ios/MobileInkExample.xcworkspace -scheme MobileInkExample -configuration Debug -sdk iphonesimulator -destination 'generic/platform=iOS Simulator' CODE_SIGNING_ALLOWED=NO build
• Simplify review: no remaining maintainability findings after cleanup.
• Security review: no auth, isolation, injection, secret, or sensitive-data findings.
• iPad benchmark evidence captured for CPU vs Ganesh manual notebook recording and 12-page suite.