How does Architecture Influence the Base Capabilities of Pre-trained Language Models? A Case Study Based on FFN-Wider Transformer Models