feat: power & thermal benchmark infrastructure

[john-rocky]

Summary

Adds mJ/token energy measurement, 30s thermal trajectory sampling, and CSV export to the sample app's Bench tool — plus the methodology doc and device-agnostic runbook.

Code (Examples/CoreMLLLMChat)

• BenchmarkResult gains thermalTrajectory, mJPerToken, timeToFair, timeToSerious, drainedPerHour, batteryCapacityWh (iPhone 17 Pro default 14.03 Wh, override per-device), and a csv() exporter with per-30s rows + # summary block.
• Bench menu adds power-reporting presets: 2 min (speed), 15 min (power), 30 min, 60 min. 15 min is the minimum that gives a defensible mJ/tok given the gauge's 1 % resolution.
• Result summary now includes Energy/token, Time→fair, Time→serious, CSV filename, and the per-30s thermal trajectory. CSV auto-saved to Documents/bench-<unix_ts>.csv.

Docs

• docs/POWER_BENCHMARK_PLAN.md — metric tiers, test matrix, head-to-head protocol against other on-device LLM engines
• docs/POWER_BENCH_RUNBOOK.md — shareable guide for any iPhone 15 Pro+
• docs/BENCHMARKING.md — new mJ/tok / %/hr / thermal-trajectory section with the API reference. Legacy J/tok derivation kept under a "legacy derivation" header.

Honest caveat documented in both docs: mJ/tok comes from iOS's 1 % battery gauge, not a lab power meter — trust only for runs ≥ 10 min.

Extracted from feature/power-benchmark (commit 78def26). The README Power & Thermal table and the cross-vocab drafter commit (2f3010c) on that branch are intentionally left out — README has been reorganized since, and the table values were _pending_ placeholders.

Test plan

• Build Examples/CoreMLLLMChat in Release on iPhone (15 Pro / 17 Pro)
• Run Bench → 2 min (speed) — sanity check tok/s + CSV save
• Run Bench → 15 min (power) — verify mJ/tok is non-zero and the trajectory has ~30 entries
• Confirm CSV is retrievable via Files app under the CoreMLLLMChat Documents directory
• Verify summary shows Time→fair / Time→serious correctly when device is cold (never) vs warm