Pure binary embeddings with MRL + hybrid PyTorch approach

Uses binary embeds with MRL, no rescoring and a new architecture for CTX files (renamed from omom), logging is rebuilt, for more stability. Removes usage of llama.cpp due to memory leaking.