I don't think these graphs showing your polling stability on desktop with nothing running are useful data. When you play a game, you will start getting tons of interrupts as the same thread that handles polling will also start handling game logic. If you play a game with mousetester recording active and then observe afterwards, it's not going to look as tranquil as your desktop graphs.
The real way to maximize polling stability in games should be:
1) identify the thread which controls USB polling by moving your mouse and observing thread activity. For me it's thread #2, or "thread #1" if you start from 0 like HWINFO64 does
2) move all possible computation off of that thread ie. forbid the game from using it
Also it should obviously help to have a stronger per-core performance, since the CPU will handle interrupts that much faster.