bilby/gw/detector/calibration.py: extrapolate calibration errors for smooth transition

This fixes #637 (closed). Since this is a corner case, I am happy to close this if extrapolation is not preferred for other reasons. This also benefits the multiband likelihood, as it requests waveform values beyond the high-frequency cutoff and assumes smoothness of waveform there, although log-likelihood errors introduced by the abrupt termination of calibration errors are typically below unity.