QFFN-BERT: An Empirical Study of Depth, Performance, and Data Efficiency in Hybrid Quantum-Classical Transformers