Sorry for the late reply on this. The various insfilters are based on a few sources. At least some of the MARGGPSFuser is based on the InertialNav library:
The other flavors of insfilter draw on that but other sources as well.
You are correct that the bias for the gyroscope and accelerometer are modeled as a Gauss Markov processes in the MARGGPSFuser. However, the gyro and accel inputs are integrated (trapezoidal) to produce a delta angle and delta velocity. These are used as control inputs to the predict phase of the EKF. So the internal EKF tracks the bias in the delta angle and delta velocity.