… __del__ against shutdown races

Closes audit items IntelPython#5 and IntelPython#29 from the prior solver review. Item IntelPython#4
(_matmat default uses a per-column matvec loop) is closed as wontfix:
SciPy's scipy.sparse.linalg.LinearOperator and cupyx's analogue both
ship the same hstack-of-matvecs default, so dpnp matches the
reference exactly and there is no portable improvement to make
without subclass-level _matmat overrides (which _CustomLinearOperator
already exposes via its matmat= constructor argument).

scipy/sparse/linalg/_interface.py
  - Set __array_ufunc__ = None on the LinearOperator base class.
    This is the SciPy contract: a host numpy.ndarray on the left of
    np_array * linop or np_array @ linop previously triggered
    NumPy's ufunc dispatch first, which would attempt to broadcast
    the operator element-wise before falling back to its reflected
    operator method -- producing either an opaque error or a wrong-
    typed result. With __array_ufunc__ = None NumPy returns
    NotImplemented from the ufunc protocol and Python's operator
    dispatch falls through cleanly to LinearOperator.__rmul__ /
    __rmatmul__. dpnp.ndarray itself sets __array_ufunc__ = None
    (see dpnp/dpnp_array.py:222) for the same reason, so the two
    dispatch systems now agree.

scipy/sparse/_csr.py, scipy/sparse/linalg/_iterative.py
  - Harden __del__ in csr_matrix and in _CachedSpMV against the
    interpreter-shutdown race where the compiled _sparse_impl
    extension is garbage-collected before the matrix instance whose
    oneMKL handle it owns. Previous code used a single
    except Exception: pass which silenced two qualitatively
    different failure modes:
      1. shutdown race -- extension gone, si._sparse_gemv_release
         evaluates to None or AttributeError; the handle is
         unrecoverable and leaving the OS to reclaim it at process
         exit is the only sane option;
      2. genuine backend error while the interpreter is healthy --
         a real bug we want to surface eventually, but raising from
         __del__ produces only an 'Exception ignored in:' warning
         and the handle is gone either way.

    The new code probes getattr(si, '_sparse_gemv_release', None)
    explicitly so case (1) takes the fast non-call path, and then
    splits the except into (AttributeError, TypeError) for case (1)-
    style residuals (queue / handle attribute access racing the
    shutdown) versus a final broad except for case (2). Both still
    return silently from __del__ -- raising is never valid here --
    but the intent is now documented and a real backend regression
    is no longer indistinguishable from the GC race in code review.

tests/third_party/cupyx/scipy_tests/sparse_tests/test_linalg.py
  - test_array_ufunc_opt_out: asserts the __array_ufunc__ = None
    marker is present on LinearOperator. Mirrors SciPy's own test
    suite test_interface.py::test_array_ufunc_opt_out.
  - test_numpy_scalar_times_linop_dispatches_to_rmul: the concrete
    runtime consequence -- numpy.float64(2.0) * linop must
    produce a scaled LinearOperator, not raise or yield an array.