fix: Add ternary weight quantization at model load time

BitNet b1.58 requires both:
- 8-bit activation quantization (per-token)
- Ternary weight quantization (per-tensor mean-based)

Now quantizes all linear projection weights to ternary values
during model loading for correct inference.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: Antigravity Agent