For question 1, the documentation goes into a little more detail about the rescaling methods here: https://www.sawtoothsoftware.com/help/analysis/manual/index.html?standard-rescaling-methods.html.
Generally we see the rescaled scores that sum to 100 used to report to others, as the probability of an item being chosen as best in kind of an average task is a little harder to explain.
For question 2, we have a chapter available from the book Becoming an Expert in Conjoint Analysis that goes into pretty good detail on stat testing choice models here: https://sawtoothsoftware.com/resources/technical-papers/statistical-testing.
But, the app itself has a short section on this as well in its help: https://www.sawtoothsoftware.com/help/analysis/manual/index.html?testing-for-differences-betwee.html
For #3, TURF is an interesting alternative to look at bundle optimization instead of just taking the top items on average, though there are a few ways to classify if someone is "reached" or not. They are discussed here: https://www.sawtoothsoftware.com/help/analysis/manual/index.html?turf.html.
My personal opinion is that the first choice is probably a bit too extreme (though still valuable), and while the weighted by probability approach isn't really strict TURF, I like it a little more. The anchor approach is the most legitimate, but requires that you asked specific anchoring questions in the survey to use.