Consolidate vision finetune files under finetune/vision/

Wire MiniCPM5 text Modal pipeline

Add Modal Stage 2 vision payload eval pipeline

Pin Modal jobs to marinaleitecabrera workspace

Fix requirements.txt: use top-level deps, not pinned sub-deps

Add both CPU and CUDA llama-cpp-python wheels, @spaces.GPU on inference

Add Viva Mais AI build spec and Phase 0 harness