Keep translation in transform_log_from_transform for tiny rotations

[jashshah999]

transform_log_from_transform returns an all-zero matrix for a transform whose rotation is tiny but not exactly identity, dropping the translation:

import pytransform3d.rotations as pr, pytransform3d.transformations as pt
A2B = pt.transform_from(pr.matrix_from_axis_angle([0, 0, 1, 1e-8]), [1.0, 2.0, 3.0])
pt.transform_log_from_transform(A2B)  # all zeros
pt.transform_from_transform_log(pt.transform_log_from_transform(A2B))  # identity

norm(I - R) is about 1.4e-8 here, so the early pure-translation branch is skipped, but compact_axis_angle_from_matrix returns a zero vector, and the theta == 0 branch then returns the log without p. exponential_coordinates_from_transform has the same two branches and already keeps p in the second one; this does the same.

test_transform_log_from_almost_identity_transform expected all zeros for an input whose translation is about 1e-4, so it was asserting the dropped translation. I changed it to check that the rotation part is zero and the translation is p, and added a test for the case above.

Found with a randomized round-trip check over rotations near 0 and pi. Full test suite passes locally (3095 passed, 3 skipped for open3d).

[AlexanderFabisch]

Thanks again @jashshah999 . I need a bit more time to analyze this though.

[AlexanderFabisch]

Hi @jashshah999 , one problem here is that the angle calculation of axis_angle_from_matrix is not precise enough. Could you rebase on #376, which improves this? Then we should be able to delete the second if entirely:

-    theta = np.linalg.norm(omega_theta)
-
-    if theta == 0:
-        transform_log[:3, 3] = p
-        return transform_log

[jashshah999]

Rebased on #376 and removed the theta == 0 branch. With #376 the tiny-rotation case goes through left_jacobian_SO3_inv correctly: round trips over angles from 0 (incl. 1e-16) to pi - 1e-6 with random translations stay within 8e-10.

One thing I noticed while testing: #376 on its own fails test_axis_angles_from_matrices_pi_general_axis (the batch axis comes back negated at pi) and test_transform_log_from_almost_identity_transform (the second is the test this PR updates). That's also what CI on #376 shows. Only the batch one remains on this branch.

[AlexanderFabisch]

Thanks a lot for this contribution @jashshah999 . With #378 , all tests pass. The PR is on develop now.