feat(hslm): TTQ — Trained Ternary Quantization with learned thresholds per layer

[gHashTag]

Summary

Trained Ternary Quantization (TTQ) — learn per-layer thresholds instead of fixed quantization boundaries.

New file

• src/b2t/ttq.zig — 196 LOC

Key insight

Standard ternary quantization uses a fixed threshold (e.g., |w| > 0.5 → ±1). TTQ learns the optimal threshold per layer via STE gradient, improving accuracy by adapting quantization boundaries to each layer's weight distribution.

Features

• TTQLayer: per-layer learned threshold with STE gradient
• quantize(): weight → {P, Z, N} based on learned threshold
• quantizeScaled(): scale-dependent threshold for progressive quantization
• computeGradient(): straight-through estimator for threshold update
• sparsity(): fraction of zeros at current threshold
• effectiveBits(): entropy-based effective bit width
• TTQNetwork: multi-layer threshold management

Tests (6)

• Basic quantization, threshold update, sparsity, scaled quantize, multi-layer, effective bits

Closes #320