Fix Raman invariant preparation

[mesonepigreco]

Summary

• centralize polarized, mixed, and seven-channel unpolarized Raman vector construction
• prepare each real-space and Q-space Raman perturbation in one backend-specific projection call
• compute add=True perturbation moduli from the complete accumulated projected vector
• document the equivalent normalized and raw unpolarized Raman conventions and correct the required channel weights
• add fast regression coverage for Raman invariants, backend parity, modulus accounting, interpolation inheritance, and IR scaling

Root cause

DynamicalLanczos.prepare_raman(unpolarized=...) accumulated diagonal components for channels 0–3, but an unconditional final preparation call overwrote the result with the last Cartesian component. The real-space and Q-space add=True paths also stored the norm of only the latest projected component rather than the accumulated vector.

User impact

Unpolarized Raman channels 0–3 now contain the complete rotational-invariant combinations. Previously saved channels 0–3 must be recomputed; channels 4–6 remain reusable. Polarized Raman and IR formulas and scaling are unchanged.

Validation

• focused Raman/IR tests: 11 passed, 1 interpolation-only skip
• normal real-space/Julia tests: 10 passed, 3 skipped
• full Q-space tests: 31 passed, 8 skipped
• Python syntax compilation passed
• MkDocs build passed

The strict MkDocs build still reports the repository's existing API-docstring warnings; the standard documentation build succeeds.