The latency numbers don't surprise me too much. With a solid-body model, you're exciting pickup and guitar structural models using a real string. When you do something other than a solid-body, you're using the real string to excite a virtual string model, which then couples into the pickup and structural models. When you use alt tunings, you have to determine the freq of the real string, then artificially excite a string model, yadda yadda.....
I suspect the identification of the real string frequency is a neat trick, and involves something more interesting than just doing FFTs, which involves a whole bunch of samples to get it right and accurate.
Thinking about it, our ability to accommodate latency is pretty amazing. Consider the time latency from when a piano player decides to play a note to the time the hammer strikes the strings. The hand has to descend, finger extends, key descends, hammer lifts, hammer leaves the key and travels to the string. It's a nontrivial amount of time. It's a neat trick of anticipation to make it so the hammer hits the string right when you want it. Same with drums/percussion.