Add RaBitQ module + comprehensive 1-bit/2-bit benchmark suite

Update CLAUDE.md with llama.cpp status, PPL results, and TODOs

Credit @ParaMind2025 for PlanarQuant and IsoQuant

Update README: speed benchmarks, architecture evolution, commit history

Update README: symmetric 3-bit PPL results beat TurboQuant

Add Llama 3.1 8B benchmarks: 239 tok/s decode, PPL 8.44, 4% faster than FP16

Update README with RTX 5090 llama.cpp CUDA benchmarks

Rewrite README: concise, results-first, reproducible

Update README: delete Mac Metal numbers, add authoritative CUDA PPL

Add post-prefill PPL benchmarks: IsoQuant 4-bit 9.03, PlanarQuant 3-bit 10.12

Restore RotorQuant trivector centroids, add CUDA PPL to README

Add IsoQuant + PlanarQuant backends to PPL benchmark

Add wikitext download step to PPL benchmark instructions

Update README with PPL benchmarks: iso3 is 2.6-43x better than turbo3

Add llama.cpp Metal benchmark results to README