llama.cpp fork adds Turbo, Planar, Iso quants

This fork represents the bleeding edge of community-driven LLM optimization, prioritizing extreme memory efficiency for consumer-grade hardware.

• –Planarquant specifically targets KV cache optimization, potentially unlocking 256k context windows on mid-range GPUs.
• –Turboquant focuses on maximizing token throughput by utilizing architecture-specific GPU kernels from Turing to Ada Lovelace.
• –Isoquant balances precision and bit-width, aiming for higher accuracy than standard 4-bit quantization methods.
• –The project includes critical manual fixes for Windows/MSVC compatibility, addressing common friction points for local LLM users on Windows.
• –Currently NVIDIA-focused, it highlights the ongoing divergence between general-purpose llama.cpp and specialized hardware-optimized forks.