Instead of mixing two consecutive biquad orders with a linear xfade, the
code used the same weight for both signals to mix, due to a copy-paste
blunder. Brown-bag fix this now. Also change a little the threshold for
using 32-samples chunks to cater for steepness jumping from 0 to 4 (this
doesn't change anything for commonly used rates and is just for
robustness in never-happening cases anyway).