Chart-based Reasoning: Transferring Capabilities from LLMs to VLMs