Are you really sure you want to do this? Roll/Pitch/Yaw is known in mathematical terms as ‘Euler angles’. There are conversions from matrix to Euler on the net (googling for Euler & Matrix should find them). However, Euler angles are not very good, as there are sometimes multiple euler specifications for the same angle, and there is a situation called ‘gimbal lock’ where you lose a degree of freedom.
If the matrices that you are tweening are < 90 degrees apart in the axis (the dot products between the axis of each matrix are all positive), you can actually get away with linearly interpolating & re-orthonormalising. This is sufficient for most games. Just adding keyframes to the path to ensure this is the case is usually fairly easy, and does not add too much overhead. At runtime, this is the fastest solution.
For larger differences, you are better off looking at interpolating Quaternions. These accurately describe rotations - although a bit of care is needed to distinguish positive & negative quaternions that evaluate to the same rotation. This is computationally more expensive that a linear interpolation but accurate for rotations up to 360 degrees in any axis. Again, googling for ‘Quaternion Interpolation’ should throw up some relevant sites.
As a test (during my last job), I wrote a character animation blending system (used on a few published console games), and challenged the artists to spot which one was using linear interpolation & which one was quaternion. They could not tell the difference, which just goes to show. Remember that for a game, all you need is ‘perceived quality’ - not rigorous mathematical accuracy.
Your reply was very well though-out, if only I could understand most of it! I’m afraid we haven’t been taught anything about 3D in my Uni course yet. However, your mentioning of simply adding keyframes seems to be the trick.