I think this should work: https://www.desmos.com/calculator/trk3nrmtkp
It calculates the amount of distortion there would be at the location of the cursor at a specific fov. This will become the cursor speed multiplier, simulating the distortion of 3D.
You would need to make a script that can continuously get the cursors xy position, calculate the distortion amount and then feed that into RawAccel somehow.
Before going through the effort I think the whole thing is quite pointless if the desktop itself isn't distorted. If it's flat then there is no context to the cursor speed change. The cursor and desktop panning is already behaving correct to the desktop's fov of 0. You aren't looking at your arbitrarily chosen fov so the cursor speed changes will feel random. If you were able to distort the desktop after it has rendered then the cursor would get distorted anyway.
Edit: I did a quick test to verify that the distortion calculation was correct as i wasnt sure. It says the distortion/speed multiplier will be 2 when moving the cursor to the top or bottom of the screen (y 0 or 1080) at 90 vertical fov. So in Aim Lab's creator studio I set the fov to 90 and placed a pillar that begins at the cursor and ends at the exact top of the screen. The sphere at the original crosshair position gets distorted vertically by a factor of 2 (from 34px to only half of the sphere being 34px).
I then changed to 150 vertical fov. The pillar was now 145 pixels tall instead of 540, which makes sense since 540 / tan(150/2 *pi/180) = 144.69. I looked up 145 pixels and the sphere again stretched vertically by a factor of 2 (from ~9px to ~18px). In the calculator if you set 150 fov and cursor y to 540 +- 145, you get 2, so it appears that my calculations were correct after all.