Anouar - #39

… Now corrected. The RPA examples are still broken and need to be looked at.

GPU-accelerated counterparts to the CPU solvers, for systems/Ncheb too
large to run on CPU. Chebyshev recurrence, MPO/MPS products, Hadamard
products, projections, and real-space/QFT sampling run on GPU (NDTensors
CUDA backend, ComplexF32); setup, Tucker SVDs, k-/spatial-group
bookkeeping, and final scalar accumulation stay on CPU.

Entry points:
- get_conductivity_ward_gpu / get_conductivity_cheb2d_gpu - sigma(omega)
- get_bands_gpu - A(k,omega) band structure
- get_ldos_spatial_gpu - A(r,omega) real-space LDOS, with :point/:block
  sampling and sublattice :average/:resolve modes
- get_dos_stochastic_gpu - stochastic-trace DOS
- get_exciton_ldos_spatial_gpu - A(X,omega) exciton LDOS
- get_C_gpu - real-space Chern marker
- scf_magnetic_hubbard_gpu, get_scf_magnetization_gpu,
  get_scf_bands_gpu - collinear magnetic Hubbard SCF loop and
  post-hoc magnetization/band diagnostics

CUDA stays an optional dependency: CUDA is resolved at runtime via
Base.loaded_modules (no `using CUDA` at load time, not in Project.toml),
so including this file does not force CUDA on users who never call
*_gpu functions.

…F maps

Adds a single planner used by every spatial sampler (eval_mps_spatial here,
plus get_ldos_spatial, get_ldos_spatial_gpu, get_scf_magnetization_gpu) so
they share one geometry-aware notion of "where to sample" and "how to
reduce the cells under each pixel."

- spatial_sampling_plan(L; ...) - builds sample centers/groups for a
  2^Lx x 2^Ly system. Supports two reduction modes:
    - :point (default) - sample/box-average at grid positions; cheap but
      aliases thin features on a coarse grid.
    - :block - gap-free coarse-graining: integrate each block by tracing
      out the within-block position bits (cost independent of block size),
      for large-scale maps of thin edge/domain-wall features.
  Also decides whether multi-atom unit cells are sublattice-:resolved
  (per-atom columns) or :averaged (one value per unit cell), or :auto
  based on sampling resolution. Extensive docstring documents all three
  procedures with usage examples.

- _eval_block_mps(A, ixp, iyp, a, b, Lx, Ly) - CPU block-integration MPS
  evaluator backing reduce=:block.

- eval_mps_spatial(A; ...) - new spatial sampler for profile MPS (e.g. SCF
  density/magnetization), built on spatial_sampling_plan; same
  num_x/num_avg/x_groups/box_half API as get_ldos_spatial but evaluates via
  eval_mps directly (no KPM recursion).

spatial_sampling_plan, requires num_x/num_y powers of two, :mpo mode
only) alongside the existing :point sampling, plus sublattice=:resolve
/:average/:auto for geometry-aware sublattice handling. New num_y,
grid, xwin, ywin, box_half kwargs enable 2D grid/window sampling.
accumulate_Tn! is simplified via a new spatial_vals_cpu closure that
unifies point vs. block reduction across resolved/averaged sublattice
layouts. Docstring rewritten to document the sampling procedures,
sublattice decision logic, and resulting shapes.

get_dos_stochastic: add dos_weighting=:trace (default, unchanged
behavior) / :sample (raw sampled signal before sector-size weighting,
for inspecting exciton bound peaks) and HODC reconstruction support
via new eta/m_order kwargs. New _dos_weight_matrix helper centralizes
the per-kernel weight matrix and normalization denom, replacing the
old _kpm_weight_matrix + hardcoded pi^2*Ncheb norm, and is shared with
the new exciton LDOS path.

Exciton LDOS: replace the cache-based get_exciton_ldos (now internal
_get_exciton_ldos_cached, kept for legacy use) with
get_exciton_ldos_spatial, an online MPS Chebyshev recursion over one
or many bound-pair positions |X,X> with no cache stored on H. New
public get_exciton_ldos (scalar and vector-omega forms) wrap
get_exciton_ldos_spatial for a single position. Update the Pathway
2/4 overview at the top of the file to match.

Minor: fix "HODC" expansion (High-Order Damping Correction ->
Higher-Order Delta Chebyshev) in get_ldos_hodc_from_mun's docstring.