Currently, doing 2,000 (i.e. nRowsPerAmp) separate calls to afwMath x16 amps to take advantage of the integer median code is untenably slow. We can either find a way of dealing with the AuxTel bias effect that makes this necessary in some other way, find a way of doing the same as afw does in a quick way in numpy/numba/something-in-python, or add code to afw that allows this in a single call (it's not intrinsically slow; other afw stats methods allow a single call to return a column-length vector for overscan, just this treatment isn't supported.)