Euclid: Supercharging Multimodal LLMs with Synthetic High-Fidelity Visual Descriptions