Since there is a lot of cross-section between computer vision and computer graphics for 3D pose and shape estimation of non-rigid objects, I would like to know what "blend shape", "pose blend shape" and "shape blend shape," exactly mean in 3D modeling of human shape and pose?
Blend shapes, also known as morph targets, are a widely used character animation technique that helps get around the limitations of linear-blend skinning (skeletal animation). A blend shape is defined as a set of per-vertex offset vectors, usually in a localized region of a character mesh, such as a neck or upper arm. The idea is that during animation, the basic pose produced by skinning is supplemented by adding these offset vectors, scaled by an animated weight factor, prior to applying the skinning transformations. This allows character artists to model such motions as the bulging/thinning of muscles when a limb is extended or contracted, or the shifting of soft tissue around the underlying rigid skeleton, which are not captured by skinning.
Blend shapes are also commonly used for facial animation, with different blend shapes representing different motions of the facial muscles, such as raising/lowering eyebrows, squinting the eyes, flaring the nostrils, moving the lips in various ways, etc. These are very difficult to model realistically using linear-blend skinning alone. By manipulating the weight factors of all these blend shapes, the animator can produce various facial expressions. A captured performance of an actor can be "solved" into animation curves for a predefined set of facial blend shapes, as a means of compression, or for transferring performance to a different face.
In the most basic version of blend shapes, their weight factors are animated by hand. But their weights can also be driven programmatically by some other set of parameters—for example, by the pose of the skeleton. Essentially, you can say something like "whenever the elbow is bent, apply the bicep-bulging blend shape", so that will happen automatically based on whatever the skeletal animation is doing, without having to animate it manually. This is broadly called "pose space deformation", since the blend shapes (deformations) are associated with regions in "pose space", the high-dimensional space of all the joint rotations/translations. The slide deck you posted refers to this as "pose blend shapes", however the more standard term is "pose space deformation" (and there's plenty of literature on it).
The "shape blend shapes" in that slide deck appear to be something specific to the pipeline they're describing, for creating a parameterized human model that can output many different body shapes and proportions. The "shape" parameters in the model presumably index the range of individual variation in body shape—different heights, limb lengths, hip to waist ratios, etc. They attempt to do this by starting with a single generic human mesh and applying various blend shapes (as well as moving joints around) to parameterize this space of body shapes.