CUDA kernels for the Muon optimizer: Newton-Schulz polar iteration (quintic + Gram-NS variants), benchmarked and verified against torch.opti

This repository hasn't been mirrored yet. Its content will appear here once it syncs.