Fix pAE-related bugs

[qiaoqiaoLF]

In the paragraph 2 of section 4.3.2 of the AlphaFold3 supplementary materials, the variable $\Phi_i = (a_i, b_i, c_i)$ explicitly refers to the coordinates of the frame atoms, not the rotation matrix as previously implemented.

The prior implementation mistakenly calculated the rotation matrix for the frames based on the atom coordinates and then incorrectly set $\Phi_i$ to this rotation matrix. This led to a misalignment between the expected coordinates and the rotation matrix used during the alignment error computation.

The bug has been fixed by correctly assigning $\Phi_i$ as the coordinates of the frame atoms. Specifically:

• The atom positions for the frame are now gathered and rearranged directly as $\Phi_i$, without first converting them into rotation matrices.
• The alignment error is computed using the frame atom coordinates rather than their corresponding rotation matrices, ensuring consistency with the intended definition in the supplementary materials.

This fix ensures that the computed alignment error accurately reflects the comparison of atom positions as described in the AlphaFold3 documentation.

[lucidrains]

@qiaoqiaoLF thank you Lifeng! 🙏