Update non-LLM benchmarks post C++ port

[chmjkb]

Current benchmarks in docs are outdated. There are significant improvements for some models after porting our code to C++. This tasks aims to replace the current benchmarks so they are aligned with the performance after the port.