Introduces DiffEngineParamConeProg, which replaces the tensor-based
ParamConeProg with a C expression tree that evaluates A, b, q, d (and P
for QPs) via the diff engine's automatic differentiation. Uses the
evaluate-at-zero trick: since all expressions are affine in x after
DCP2Cone, the Jacobian gives A and the gradient gives q directly.

Also adds SymbolicQuadForm and DivExpression converters to the diff
engine, updates ConicSolver/QpSolver to accept the new program type,
and includes comprehensive comparison tests against the SCIPY backend.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Resolve canon_backend=None to DEFAULT_CANON_BACKEND in
construct_solving_chain so that changing the default actually
propagates through all code paths.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

DIFF_ENGINE doesn't yet support all atoms, so keep CPP as default.
The matrix_stuffing benchmarks now use ASV parameterized benchmarks
to compare both backends side-by-side.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

n_params was only used for a size validation check and can be computed
from n_param_nodes + each node's size. This simplifies the API by
removing one argument from problem_register_params and one field to
maintain.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

The dimensions are now read directly from param_node->d1/d2 inside the
C implementation, so the Python binding and caller no longer need to
pass them.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Precompute total parameter size when parameters are registered instead
of recomputing it on every update_params call.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Document why these use generic make_index rather than custom C nodes:
hot-path cost is dominated by Jacobian memcpy, not index lookup, and
the Hessian sparsity limitation is identical either way.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

# Conflicts:
#	CLAUDE.md
#	cvxpy/reductions/solvers/nlp_solvers/diff_engine/_bindings/bindings.c
#	cvxpy/reductions/solvers/solving_chain.py
#	diff_engine_core

The import used `import _diffengine` directly, which isn't available.
Updated to `from sparsediffpy import _sparsediffengine as _diffengine`
to match the rest of the codebase.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

Inherit from ParamProb so DiffEngineParamConeProg is no longer a
duck-typed stand-in: move split_solution to the base class, simplify
isinstance checks to use ParamProb, deduplicate format_constraints by
delegating to _build_restruct_operator, and fix PowConeND ndim==1 bug
in the shared operator builder.

Co-Authored-By: Claude Opus 4.6 <[email protected]>