Fix low-mode vibrational analysis to use mass-weighted coordinates

[mooreneural]

The GF-matrix method (Wilson, Decius & Cross) requires building the rigid-body
translation/rotation basis and projecting the Hessian in mass-weighted
Cartesian coordinates q = M^{1/2} x. The previous implementation used
unweighted Cartesian coordinates, causing heavy atoms (Br, I, S) to artificially
inflate mode eigenvalues and misrank the true soft conformational modes.

Changes

• _build_vibrational_basis: translation vectors now weighted by √m_i per atom;
  rotation vectors weighted by √m_i · (r_i × e_k); near-zero threshold scales
  with √(total_mass) accordingly.
• _select_low_modes: projects the mass-weighted Hessian F_mw = M^{-1/2} H M^{-1/2}
  into the vibrational subspace, diagonalises, then unweights eigenvectors back
  to Cartesian displacement vectors and renormalises columns.
• config.py: corrects low_mode_eigenvalue_threshold docstring units from
  kcal/mol/Ų to kcal/mol/Ų/Da (mass-weighted, proportional to ω²) with
  calibration reference values for torsional vs. stretch modes.

Benchmark

Cosine similarity measures alignment between the old and new softest-mode vectors.
Values near 1.0 mean the fix is a near no-op; lower values indicate the mode
ordering was substantially corrected for that molecule.

Molecule	Max mass (Da)	Old soft modes	New soft modes	Cosine
Propane	12	4	5	0.986
Hexane	12	5	5	0.973
Thioanisole	32	5	5	0.912
Chlorobenzene	35	4	5	0.721
Bromobenzene	80	5	5	0.680
4-Bromobenzyl alcohol	80	5	5	0.937
2-Iodothiophene	127	3	5	0.624
Dibromomethane	80	0	4	—

For C-only molecules the fix is a near no-op (cosine ~0.99). For Br/I molecules
the soft-mode vectors are substantially corrected, iodothiophene's softest mode
previously assigned 59% of displacement to the I atom; the corrected mode assigns
~3%. Mode-select overhead is <1 ms vs. 4–12 ms for Hessian computation.

[corinwagen]

Hey @mooreneural, thanks for the PR. We originally did this intentionally w/ the non-MW Hessian, following the literature precedent, but we ran our own benchmarks on this PR and performance seems good for "normal" non–heavy element systems, so I think this is a real improvement.

A couple notes:

• can you fix the errors flagged by Ruff? I think it's just because of the "x" character
• can you update the docstring for generate_low_mode_seeds to match here?

I'll be happy to merge after this changes.

[mooreneural]

Fixed both! Replaced the Unicode multiplication signs with ASCII x to clear the Ruff warnings, and updated the generate_low_mode_seeds docstring to reflect the mass-weighted units (kcal/mol/Ų/Da). Ready to merge when you are.

[corinwagen]

merged! ty for the contribution