It's not really a huge win in general.
You can use a perspective projection into a true 3D scene to get an effect like parallax scrolling background layers from classic 2D games, sure. But it's not really an easier to implement that than it is to implement scrolling layers.
You don't get any benefits with respect to depth ordering -- you'd still need to give each sprite an appropriate depth (Z coordinate) and this is no different in practice than assigning the sprite a depth layer. You get slightly smoother transitions between depth layers because they aren't discrete layers, but that's about it. You can use the depth buffer and test to get properly rendered sprites regardless of whether or not you use a perspective (versus orthographic) transform -- and either way that's a wash since you generally have to manually sort sprites back-to-front for rendering any that have alpha. You actually get that basically for free with a regular background layer index.
You can counter the resizing by pre-scaling the sprites based on depth, typically in a shader. This avoids the need to try to rely on mip-mapping to fix up the appearance (mip-mapping is a bit of a red herring here, as it only improves the visual quality of the resulting textured quad and does not fix the fundamental resizing problem and it's not why it was developed). On the other hand, you may find this scaling desirable if you're going for an obvious "2D in 3D" look, like Shadow Complex, versus an obvious 2D look like Super Mario World.
In short, like all things, it depends. If your game is meant to be a 3D game constrained to a 2D field of play, it makes sense. If it's not, and you don't need any of the small benefits (such as smooth layer transitions), it's probably not worth doing just because exactly equivalent results can be obtained with less work in another fashion.
If, on the other hand, your frameworks and toolchains are only set up for a perspective projection and cannot be changed or swapped out, and you are only doing a simple 2D game, you can certainly make it work. There's no clear answer to this and the differences between options are pretty minor.
Go with the option that involves the least work.