llama.cpp adds memory-fitting to llama-bench

Native memory-fitting logic in llama-bench is a massive win for reproducibility in local LLM performance testing.

• –-fitt (--fit-target) allows users to set specific memory or performance goals, automating the search for the optimal configuration.
• –-fitc (--fit-min-ctx) ensures benchmarks don't sacrifice too much context size to hit performance targets, which is crucial for RAG and long-context use cases.
• –These flags eliminate the need for external scripts or manual trial-and-error when optimizing for 8GB or 16GB VRAM cards.
• –Dedicated output columns for these parameters improve the transparency and reproducibility of shared benchmark data.
• –This update signals llama.cpp's push toward making complex LLM optimization accessible to a broader developer audience.