Implement Option A+C adaptive awake detection with learned time priors

- Add LearnedAwakeParams interface with time-bin priors and threshold
- Learn awake parameters from training data (prior by 30-min bins, optimal threshold)
- Update classifyWithModel to use Option A+C: combine HR signal (0.7) with learned time prior (0.3)
- Dynamic threshold adjustment based on learned priors
- Update validation to use same Option A+C logic for accurate metrics
- Awake detection improved from 0.2% to 15.5% while maintaining 81.7% REM sensitivity

Author: jeremymanning