Does one need to normalize the rotation computed by TransformTRS?

When using the glTF library, and decomposing the matrix ```{1,0,0,0,0,0.5,0,0,0,0,1,0,0,2,0,1}``` into TRS, I get a non-unit quaternion with length `0.875`. Is this intented behavior?

As a side-note, would it be interesting to use SVD to also allow shearing, by decomposing any 3x3 matrix into rotation * scale * rotation? This would mean introducing extra dummy nodes. For example [this implemention](http://pages.cs.wisc.edu/~sifakis/project_pages/svd.html) seems to be both accurate and fast.

Start at the TransformTRS implementation and reproduce the reported decomposition using the matrix in the issue. Determine whether the non-unit quaternion is expected, then define the needed behavior for normalization or correction; the SVD/shearing proposal is a separate design question.

Written by the indexing model from the issue text.