Stencyl (or any game engine, really) provides a "one size fits all" approach to memory management, but it isn't optimized for any particular kind of game. At some point you might have to take more control of asset loading/unloading to see any gains in speed.
For example, suppose you have a platformer and a player moving right. You can load images used only in the next screen as the player moves through a certain point--thus, smoothing out the loading. Note--If you load and unload assets too frequently, you'll start to run into the "opposite" problem: too much garbage collection activity. Also, the additional complexity of the programming might even give you worse performance, or be too bulky to manage.
I think a better place to start would be to make sure you've configured your atlases as efficiently as possible.
I don't know the answer to this question, but I think it'd be worth exploring: how much of the load time is spent processing images? All of the images are stored as .png files, which have to be decompressed. Would replacing the .png files that Stencyl generates with files saved with less compression improve load times? If you have ImageMagick, you could run a batch recompression on all of the .png files. This would increase the file size of course, but disk storage on PCs is cheap so it may be a worthwhile trade-off (assuming this "experiment" works at all, that is)