These algorithms are fantastic for microcontroller applications. You can even choose the multiplicative constant so that it fits in a Thumb-2 immediate and save yourself a constant pool reference, with no degradation in quality for the output (as measured by spectral tests).
It's true that needing strict unpredictability, but not true cryptographic security, is rare in a microcontroller application. Though it can be nice to have.
What makes PCG attractive for microcontrollers is really that it's of known good quality, its implementation is very small and very simple, and it ends up generating efficient code for 32-bit processors (i.e., ARM Cortex-Ms). That is not, and can not ever be, true for something like a 128-bit shift register. PCG is great for just tossing in when I'm working on a platform that has no built-in library rand() function, or where it's busted, or whether I don't want to bother figure out if/how badly it's busted. (With how easy PCG is to use, that last one covers every embedded platform ever....)
PCG is not the best RNG out there: it's not the highest quality, it's not the fastest, it's not the least predictable, it's not the strongest theoretically. (It isn't the smallest, either, but it's quite small and the smaller RNGs I know of are either code-golfed, which doesn't count, or truly garbage.) But it is a nice equilibrium between all of those things. And did I mention it's simple? Simple is really, really nice to have :)