I think we wouldn't notice the difference. The difficult tasks are probably already there, but the LGMs can make one feel like they were really easy. Also, such a round would just shift the normal, as the top finishers would still anyway compete at break neck pace, as it's still a race. Besides, i guess inventing such harder problems is not an easy task. It takes time and expertise that only a few possess.

Time is directly proportional to cheating. As time increases cheating also increases. So, 2-hour duration is the best, I agree here also cheating is done but when compared it is least.

IT HAS NO END. https://en.wikipedia.org/wiki/Imperial_examination

CodeChef Starters contests (for example this) and Google KickStarts (for example this) have 3 hours duration. They indeed allow slower participants to have a realistic shot at solving higher difficulty problems during the contest time.

Additionally I would recommend AtCoder heuristic contests (for example this). That's 4 hours to solve just 1 problem. Though heuristic problems differ from what you can see in a typical algorithmic contest and these heuristic problems are much closer to what real software developers do as their everyday job.

This is a nice idea

It's worth mentioning that if you have this, you need to make at least the first problem super easy. Currently there's an interesting feature of the rating system which is that people who solve no problems don't lose elo, which means if you solve just the easiest problem with bad penalty, you'll be at the bottom of the scoreboard and lose elo, despite beating everyone who solved 0 problems.

So to avoid this, there needs to be at least one simple problem everyone can solve to see who is really participating in the contest.