Allow aligning the mag to any rotation

[fiam]

Introduce 3 new variables which allow setting the decidegrees
for the mag sensor alignment. When any of these 3 variables
are non-zero, mag is assumed to be mounted off-board and
"align_mag" as well as the board alignment are ignored.

Settings are named align_mag_roll, align_mag_pitch and
align_mag_yaw.

Fixes #86
Fixes #1029

[fiam]

Closes #2014

[digitalentity]

rotateV is quite expensive because it builds rotation matrix every time.

[fiam]

@digitalentity Just moved the matrix computation to compass initialization. I'm wondering if we should do this also for on-board mag alignment. The 9 mul + 6 add used by the rotation might be faster than the current branching + muls used by applySensorAlignment() + applyBoardAlignment(). I'll profile it later.

[fiam]

Just profiled this. First graph is on-board alignment with no board rotation, second is on-board alignment with some board rotation and third is the new code for off-board alignment:

On-board alignment, no board rotation

On-board alignment, board rotated

Off-board alignment

With a non-rotated board, performance of both approaches is basically the same. However, once board rotation is non-zero, the off-board alignment is faster. Upon reviewing the code for applySensorAlignment() and applyBoardAlignment(), it becomes clear that this happens because applyBoardAlignment() returns early when board has no rotation but performs a multiplication by a pre-calculated matrix when it's not, so it does as much as off-board alignment just in applyBoardAlignment() and then we still need to add applySensorAlignment() on top. There's, however, a drawback for the additional pre-computed matrix and it's that it takes 36 bytes of RAM.

With this said, and given that right now it doesn't make much sense to have an off-board aligned acc nor gyro, I think we should leave the on-board alignment code in since there's no good reason to spend 36 bytes of RAM per sensor to have a single matrix for each (specially on F3 targets where memory is already close to the limit).