Install torchvision CPU-only in Dockerfile to match CPU torch and prevent nms operator failure

Add librosa, soundfile, and libsndfile1 for Gemma 4 audio processor dependency

Upgrade python base image to 3.10 to support transformers >= 5.5.0

Migrate architecture from Ollama to native Hugging Face transformers with CPU optimization

Add zstd dependency required by Ollama installer

Fix Ollama installation by using official script

Install Ollama and run gemma4:e2b with FastAPI wrapper

Add application files