Residual "noise" in rank(B)<3 results

[filip-szczepankiewicz]

Requests for B-tensors that are not full rank have residual "noise" on axes that do not contribute to the final b-tensor. IT would be nice to set these to zero so that they do not cause distractions. This is safe from the motion-comp, and Maxwell-comp perspectives.

[jsjol]

I thought it would be possible to fix this by making sure the initial guess was zero for the deactivated axes (because the gradients should be zero). But apparently I was mistaken, as I found out after implementing #36.

The most efficient solution would definitely be to reduce the optimization problem by excluding the deactived axes entirely, but this would require changes all over the code. Since the issue doesn't seem to be the most pressing one I'm thinking of closing this issue without further action. Do you have a better idea?