Radu's suggestion (just pull Z out of the point cloud) is very simple. The only complication is that sensor_msgs/PointCloud doesn't store the points in 2d image layout, so you'll also have to look up the pixel coordinates in the "channels" field.
Another way is to start from the disparity image, which is inversely related to the depth map. The comments to
stereo_msgs/DisparityImage explain the math. Given disparity d, focal length f and baseline T, the depth is Z = fT/d. You also need to check whether d < min_disparity, which indicates an invalid disparity and hence unknown depth.
Patrick