refactor: unify CuPy/buffer-protocol path for array and sparse Csr

[rho-novatron]

PR-A: refactor: unify CuPy/buffer-protocol path for array and sparse Csr

Branch (fork): rho-novatron:refactor/array-helper
Base (upstream): Helmholtz-AI-Energy:main
Commits: 1 (17b586d)

---

Summary

Pulls the __cuda_array_interface__ / Python-buffer-protocol input
handling out of array.cpp and matrix.cpp and into a single helper
in utils.hpp:

template <typename T>
gko::array<T> gko_array_from_pyobject(
    std::shared_ptr<const gko::Executor> exec,
    py::object obj);

Both the gko::array(Executor, py::object) constructor and the
sparse Csr(Executor, scipy/cupyx csr_matrix) constructor now go
through it.

Why this isn't just a refactor

The pre-existing sparse-matrix constructor took two structurally
different paths depending on input type:

• CuPy CSR on a CudaExecutor — zero-copy via
  __cuda_array_interface__ (added in Add CuPy zero-copy interoperability via __cuda_array_interface__ #98).
• Anything else (NumPy / scipy / buffer protocol) — a manual
  fallback that built gko::array<...>::view(...) on a
  ReferenceExecutor and handed those host arrays to
  Csr::create(exec, ...), forcing a second host→device copy when
  the executor was CUDA.

After this PR, both paths flow through gko_array_from_pyobject,
which:

• honors the actual target exec for the buffer-protocol path
  (constructs the array directly on the requested executor instead
  of via a Reference detour),
• applies the dtype typestr validation that previously only the
  CUDA branch performed, to the host path as well, and
• gives array and Csr a single source of truth, so future fixes
  (e.g. the cross-platform integer-dtype-char fix in the
  follow-up MPI PR) automatically apply to both.

Behavior change for sparse matrices on CUDA built from
NumPy/scipy input: now allocated and copied directly on the target
device executor instead of through a host ReferenceExecutor
intermediate.

Diff

 src/cpp_bindings/array.cpp  | 70 +-----
 src/cpp_bindings/matrix.cpp | 84 +---------
 src/cpp_bindings/utils.hpp  | 66 +++++++
 3 files changed, 75 insertions(+), 147 deletions(-)

Why this is its own PR

• Stands alone as a behavior-improving cleanup of Add CuPy zero-copy interoperability via __cuda_array_interface__ #98's wake.
• Reviewers can sign off on it independent of the much larger MPI
  feature work that follows in the stacked PR-B.
• PR-B's Windows-portability fix is a one-line change against this
  helper — much easier to review when the helper itself has already
  been reviewed in isolation.

Testing

The existing test suite covers all the input shapes: NumPy/scipy CSR
on ReferenceExecutor, CuPy CSR zero-copy on CudaExecutor, and
the dense-array constructors. No regressions observed locally.

Stacked PR

Not stacked on anything (based on main). PR-B for the distributed
MPI bindings is stacked on top of this.