feat(bench): BENCH-001 format comparison (MSE + dynamic range + LUT cost)

[gHashTag]

Summary

BENCH-001 partial implementation — systematic format comparison benchmark.

What it covers

1. Quantization MSE: ternary vs fp16 vs bfloat16 vs GF16 (1000 Normal(0,1) samples, seed=42)
2. Dynamic range table: min/max representable values per format
3. LUT cost estimate: ADD/MUL LUT counts on XC7A100T

What remains (needs hardware)

• Inference benchmark: same model, f16 vs ternary
• PPL comparison: f16 baseline vs ternary HSLM
• Power/throughput on FPGA

Usage

zig run benchmarks/format_comparison.zig

Refs #479