After reading the GTSAM code about Imu factor like ImuFactor.cpp, ManifoldPreintegration.cpp, etc. I have got the following conclusion:
(1) In optimization, when bias changes, we can update the delta measurments using a first-order expansion like in [1] formula(44)
(2) After optimization, we can use new bias and recompute the preintegrated imu measurements if bias changes a lot especially in initilaization.
[1] Forster C, Carlone L, Dellaert F, et al. On-Manifold Preintegration for Real-Time Visual--Inertial Odometry[J]. IEEE Transactions on Robotics, 2016, 33(1): 1-21.