BAQET: BRAM-aware Quantization for Efficient Transformer Inference via Stream-based Architecture on an FPGA