Phonons: q-only supercell diagonalization, in-place sort, LO-TO and AdjustQStar fixes

[SorBalda]

Phonons: q-only supercell diagonalization, in-place sort, LO-TO and AdjustQStar fixes

Part of a three-repository series that makes the q-space TDSCHA path linear in
memory instead of quadratic. This one carries the primitive the other two build
on; it is useful on its own for the two bug fixes at the end.

• CellConstructor: this PR
• python-sscha: Ensemble: opt-in linear-memory q-space mode, plus four fixes on the Fourier path python-sscha#428
• tdscha: QSpace: fix the Schur fill between repeated irreps, and make the q-space path linear in memory tdscha#35

All four changes are in cellconstructor/Phonons.py, and all but the last are
in the fast DiagonalizeSupercell.

1. New kwarg q_only=False

When True, the (3N, 3N) supercell polarization matrix is never allocated.
Mode selection is unchanged; for each accepted mode only its origin (iq, band)
is recorded, in two O(3N) integer arrays, so that
w_q[mode_band[k], mode_iq[k]] == w_mu[k]. Callers that only need the q-space
eigenmodes then run in memory linear in the supercell size.

This path also skips the pre-computed phase table, which is (3N, nq) and would
be the largest allocation left; the phase is recomputed one q at a time. That is
a gemv against a gemm, so the two agree to BLAS rounding, and w_q/pols_q do
not depend on it at all.

Verified on a 27-q dyn: w_mu, w_q and pols_q are bitwise identical to
return_qmodes=True, and the (iq, band) map is exact. Also checked on
non-cubic grids ([4,3,3], [4,4,3]) and before/after AdjustQStar.

2. Sort the polarization columns in place

e_pols_sc[:, sort_mask] is fancy indexing, so the sorted copy is built while
the original is still alive, doubling the peak of the routine's largest
allocation. Replaced by a cycle-following permutation using a single length-3N
temporary column.

Verified bitwise equal to the fancy-indexed result over 200 random permutations
and on real dyns with heavily degenerate spectra (up to 1541 degenerate modes
out of 1920).

3. Honour the LO-TO warning

With lo_to_split set but no effective charges, the code warned
"LO-TO ignored" and then ran the branch anyway. Since
ForceTensor.Tensor2.Interpolate guards the whole non-analytic term behind
effective_charges is not None, that branch rebuilt the dynamical matrix
DyagDinQ already has — after allocating a full supercell Tensor2 through
SetupFromPhonons → GetRealSpaceFC. Measured: 132 MB at an 8x8x8 supercell,
growing as N².

Falling through to DyagDinQ gives the same frequencies and keeps the Gamma
polarization gauge consistent with the other q points. The lo_to_split string
validation is moved above the check, so a typo still raises when the charges are
missing. Applied to both copies of the routine.

4. AdjustQStar swapped the wrong block

The search loop had no break, so iq kept its final value and the swap used
the last q visited instead of the Gamma one: the reordering was corrupted
whenever Gamma was neither first nor last. It also proceeded silently when no
Gamma point was present, in which case the swap zeroed q_tot[0] and fabricated
a Gamma that did not exist. It now raises.

---

Testing. No test data is added here. The claims above were checked against
gold and LaAlO3 dyns outside the repo; happy to add a small committed case if
you would like one.

[mesonepigreco]

Here needs a better comment explaining what the following code does rather than saying what was bad about the old code

[mesonepigreco]

Ok, just a more explicative comment in the last part of the edit (which is really incomprensible code).

Moreover, to comply with the update in the tdscha, move the version naming to 1.7.0. We need to do it anyway. (This needs to be done both in meson and in the python toml settings.