Dynamic Quantiser - a way to make your own high quality dynamic quants

The Dynamic Quantiser is a tool for creating high-quality dynamic quantizations. It precomputes dot products, squared norms, and byte counts for each tensor and quant level. The objective is to minimize whole-model cosine deviation, calculated as 1 - S / sqrt(AQ), subject to a file-size budget. This optimization uses Dinkelbach's algorithm, a separable Lagrangian inner loop, and a multi-start local search with single-tier and pair-swap moves.