Fix and modernize the dormant GPU/cuSolver build

[Prasanna163]

What

The optional GPU code path (WITH_GPU / cuSolver, originally added in 2020) had
bit-rotted and no longer configures or compiles. This PR restores the build
plumbing without touching the default CPU build — all GPU bits stay guarded.

Changes

• src/mctc/lapack/eigensolve.F90 — fix an invalid local declaration
  real(dp) :: dummy(:) used for the cuSolver bufferSize query (a local
  assumed-shape array without allocatable/dummy status is not valid Fortran and
  broke the USE_CUSOLVER build). Replaced with dummy(1), a valid placeholder
  for a size-only query.
• meson/meson.build — modernize deprecated NVIDIA HPC SDK flags for current
  nvfortran: -ta=tesla:ccNN → -gpu=ccNN -acc, -Mcudalib → -cudalib.
• CMakeLists.txt — add WITH_GPU / GPU_ARCH / WITH_CUSOLVER options
  (CMake had no GPU support at all), mirroring the meson logic; Fortran-scoped
  flags; defines XTB_GPU / USE_CUSOLVER / USE_CUBLAS.

Testing

• The default CPU build (meson + gfortran) and the unit suite are unaffected —
  every change is behind the (off-by-default) GPU options.
• The cuSolver eigensolver path these flags enable was validated standalone on an
  RTX 3050: cusolverDnDsygvd vs LAPACK dsygvd agree to ~1e-14 eV on real
  GFN0 H/S matrices.

Heads-up for context: a full GPU build of xtb is currently blocked further
upstream by nvfortran internal compiler errors on the modern-Fortran
dependency tree (deferred-length character handling, first seen in toml-f).
This PR is just the build-system groundwork and is independent of that issue —
the default build is unchanged either way. Happy to adjust anything.