Full ternary: rebuild modular detectors and drop pattern_recognition

Add --ternary flag to quantize.py; rewrite buffer gates as +/-1 weights

Quantize weights to per-tensor minimum signed integer dtype