Q8 result: thrashes on 16 GB (CPU_REPACK doubles memory)

Add Gemma 4 benchmark results

Update MLX sniper link to 5.4 tok/s

Update MLX sniper link: 5.2 tok/s on 35B

Add full README with verified benchmarks and research findings

Add common/arg.cpp

Add common/common.h

Add common/common.cpp

Add src/CMakeLists.txt

Add src/llama-expert-cache.h

Add src/llama-expert-cache.cpp

Add src/llama-expert-cache-ctx.h

Add src/llama-expert-cache-ctx.cpp

initial commit