🔄 Fix: remove --only-binary (causes cache miss), let pip prefer wheels

🔥 Fix OOMKilled: --only-binary :all: + llama-cpp-python==0.2.90 (verified wheel)

🔥 Fix: remove --only-binary (breaks build), use flexible llama-cpp-python version

Fix: OOMKilled → pre-built wheels only, 32K context, multimodal Gemma-4-E4B