computeCanonicalTransform may generate non-canonical coords

[ptosco]

This is related to #1908, #1912 where it was first noticed that computeCanonicalTransform, which uses an approximate built-in eigen solver, may not always return fully converged eigenvalues/eigenvectors of the covariance matrix of 3D coordinates.
I have noticed a similar issue while canonicalizing a large set of 2D conformations: some of the molecules were either rotated in a non-canonical orientation or applied a rotation on the X or Y axis resulting in non-zero Z coordinates.

Here's an example of what I mean; the Canonical_numpy_mol column reports the canonical coordinates obtained by implementing the canonicalization algorithm using numpy, whereas the Canonical_rdkit_mol reports the canonical coordinates obtained using computeCanonicalTransform,. The 2D coordinates appear distorted in 2D, but actually it's because the Z coordinate is non-zero for some of the molecules: