llama.cpp lands attn-rot quantization boost

This is the kind of pragmatic infrastructure improvement that matters more than flashy algorithm names: a simple baseline that preserves quality while making cache quantization safer.

• –The implementation is intentionally minimal and backend-agnostic, so it fits llama.cpp’s broad hardware story without introducing new tensor types
• –Benchmarks in the PR show q8_0 essentially matching F16 on several models, while lower-bit caches also improve versus the pre-rotation baseline
• –It is not full TurboQuant; the author explicitly notes missing pieces like PolarQuant and QJL, so this is a strong baseline rather than the end state
• –For long-context users, the practical win is cheaper KV cache memory with less quality regression, which is the main pain point these tricks are trying to solve
• –MLA is not supported yet, so the current upside is strongest for the model families llama.cpp already handles well