I don't think the complementary filter is going to cut it, in this case. Where I see the term used (and it's a new one to me -- it seems to have been invented by the smart-phone software community) it's assuming that the phone is being accelerated mildly and rotated wildly. The airplane case is almost, but not quite, opposite.
You'll need to know a lot more than just linear algebra to get your head wrapped around Kalman filtering, and then, to confuse things further, this problem actually demands an extended Kalman filter. Trust me, the math will get wacky, and demanding, and a lot of other things only some of which can be stated in a family-friendly forum.
The way that a Kalman filter works is that it sorts out redundant data. This means that in order for a Kalman filter to do you any good, you have to have redundant data to feed it. If you just have a free body moving in space (like an RC plane), then a six-axis IMU (accelerometer and gyro) does not give you redundant information: the plane can move in six axes, you have six axes, too bad. You can use a six-axis IMU plus some absolute position reference (like GPS) to get your redundancy, but in our case you'd need to add a GPS to your plane.
What saves us here is that you do have redundant information, but it is indirect and screwy: the pilot constrains the airplane to operate in a hemispherical shell with a radius that's more or less limited to the line length, and the acceleration of gravity is a constant vector of known magnitude. With that information I think that there's enough to get the airplane's position, velocity and attitude with respect to the pilot and to ground, but with a North reference that's "unpinned" and will spin around the circle at some rate that's (hopefully!) much less than the speed of flight.
This "pilot pins the center of the circle" is what you're already using to deduce speed from rotation rate, by the way. What a Kalman filter will do for you is to take into account all of the rotations and accelerations of the airplane such that (for instance) any rocking of the airplane in yaw, or any coupling between the gyro's yaw axis and the airframe's pitch axis, will be washed out of the final velocity estimate. In addition, I think (everything I say about this subject at this point should be prefaced with "I think" or "if I'm right") that as a bonus you'll know when the airplane is pointed up and when it's pointed down, and be able to add that to the motor speed as a feed-forward, to get a quicker reaction than you could from the loop.