The app. note shows a way of reducing the response time of the opto isolator, by making sure that the phototransistor is not saturated. The cascode arrangement with the BJT does that. All transistors take time to come out of saturation.
The idea of the current source is that the error amplifier stays in its linear region, even if there is a large error. That will also speed up the response when the error comes back towards zero. The data sheet points out the improvement during start-up when the error is huge and I guess that overshoot is reduced.
TL431 circuits are by far the most common for SMPS. They work fine for most applications. However, if there is a large error, the op-amp within the TL431 will be in a non-linear part of its operating range, and the output transistor within the TL431 will be saturated. Either or both of these could introduce a significant delay and result in poor transient response, because the LED will take time to turn off after the error reduces to zero. All sorts of factors will affect how bad the problem is and how much it affects the load.