Rank-reduced Woodbury kernel for the lattice Green's function

[harrisonlabollita]

Incremental diff:
pr/obe-printing-h5...pr/woodbury-kernel

Note that this PR shares postprocess.{cpp,hpp} with PR #15 for unrelated
reasons. So we should merge #15 and rebase before merging this PR.

Summary

Every k-summed quantity in ModEST — gloc, the density, the chemical-potential
search, the charge-density correction — inverted the full
(N_\nu \times N_\nu) band-basis matrix
(G_k(\omega) = [\omega + i\delta + \mu - H(k) - P^\dagger \Sigma(\omega) P]^{-1})
at every ((k, \sigma, \omega)). This inversion is unnecessary, because for the fixed k-grid cases (to which
this port applies), (H(k)) is diagonal in the band basis and (\Sigma) is structurally space in the C-space.

This PR applies the Woodbury identity to collapse the inversion. We build up the basic backend primitives that allows us to port all lattice Green's function computations onto one shared Woodbury kernel.

Changes (organized by commit)

02957d3 — extract lattice_gf_helpers.hpp (pure refactor, no numerics change
except as noted). Pulls the direct (N_\nu \times N_\nu) inversion helpers out of
gloc_fixed_grid.hpp into a new shared header so density, chargedensity and
the post-processing routines can use them:

• upfold_self_energy_at_freq, local_gf_at_k, lattice_gf_at_k move over, losing
  their Mesh template parameter.
• upfold_self_energy_all_freq moves out of postprocess.cpp.
• new make_sigma_total, folding Sigma_static into Sigma_dynamic once per call
  rather than at every frequency.
• removes dead code: the old spectral_function sketch in postprocess.hpp, the
  impurity_levels stub in downfolding.cpp.

One deliberate numerical change here: density_for_matrix_valued_impl now reuses
lattice_gf_at_k and computes trace(G_band) - trace(inv(w + mu - H_k)).

ca59384 — the kernel, and gloc ported. Adds active_subspace_t,
detect_active_subspace, compute_bare_projected, compute_sigma_active,
make_woodbury_setup, apply_K, build_G_band to lattice_gf_helpers.hpp;
gloc_fixed_grid.hpp uses them.

fb552e2 — density and chemical-potential search. trace_G_B_m_G_KS is
replaced by the rank-reduced correction (\mathrm{tr}(K \cdot G^0_{QQ,\mathrm{sq}})),
accumulated per ((k,\sigma,\omega)). The Kohn-Sham term is unchanged, and the
(\Sigma = 0) case now short-circuits to it. Removes detail::G0_C_k_sigma
(downfolding.{hpp,cpp}) and detail::calc_inv_G_G0 (density.hpp).

1b76bf1 — charge_density_correction. Builds (G_k) from the identity
instead of inverting at every frequency. Two changes here are independent of the
Woodbury port and worth reviewing separately:

• adds an explicit matrix_valued (H(k)) fallback branch. The rank reduction
  assumes diagonal (H(k)); previously every case went through lattice_gf_at_k,
  so the matrix-valued path now keeps the direct inversion.
• parallelises the ((k,\sigma)) loop with OpenMP.

Testing

This is a behavior-preserving change. No new tests
are added, because the existing reference-data suite already covers the kernel
densely and passes unchanged (including the LFS set under
-DEnable_LFS_Tests=ON).

Review notes

• Read lattice_gf_helpers.hpp first; the block comment at the
  Rank-reduced Woodbury banner states the identity the rest of the header
  implements.
• G0_QPdag is not dagger(G0_PQ) — the (i\delta) makes (D) complex. Same
  reason Lmat · KR is not Rmat† · … in build_G_band. These are the two easy
  places to introduce a wrong conjugate.
• LaTeX notes of the implementation are also available on the ModEST paper draft.