Does dcm2niix generate the correct solution for converting gradients specified in PRS convention to FSL format? I would expect that Siemens reports the directions in the DICOM as scanner based XYZ, regardless of how they are specified in the sequence. Therefore, my assumption is that dcm2niix should convert your images correctly. If this is not the case, please generate an issue on the dcm2niix GitHub page. 

By the way, I would expect the conversion of PRS to XYZ will not always match your illustration. I would expect choosing Axial, Coronal or Sagittal slice geometry, as well as phase encoding dimension and polarity will impact this transformation.