The doxygen documentation for calcBodyAccelerationFromUDot() says for parameter A_GB:
Spatial accelerations of all the body frames measured and expressed in the Ground frame.
So it is
measuring the acceleration of the body frames with respect to Ground, and
expressing the resulting pair of vectors in the Ground basis. The linear acceleration is of the body frame origin; angular acceleration doesn't depend on a point.
Could you explain why q_dot != u?
The state variables are the generalized coordinates q and generalized speeds u. Those can be chosen independently as long as there is an instantaneously-linear relationship between qdot and u: qdot=N(q)*u. Generalized speeds u are physical in the sense that their dimensionality is the same as the number of degrees of freedom in the multibody tree. Generalized coordinates q are commonly chosen instead for numerical stability or convenience. Most commonly, a quaternion (four q's) is chosen to represent unrestricted orientation, even though there are only 3 dofs and hence 3 u's. In that case there are four qdots and three u's so they clearly can't be the same thing! But there is a simple 4x3 matrix N such that qdot=N(q)*u when q is a quaternion.
If you don't have numerical concerns regarding orientation, it is no problem in Simbody to construct systems in which qdot=u; i.e. N is an identity matrix. Simple joints like pins, sliders, and u-joints already have that property, and the Gimbal joint can be used instead of a Ball to provide arbitrary orientations with 3 q's and 3 u's=qdots. But that necessarily has a singular configuration.
You might find
this paper helpful in understanding the parameterization of multibody systems in Simbody, but also any basic discussion of rigid body mechanics will include at least quaternions and thus have the qdot != u problem to deal with.
Sherm