Extract PSD cones from unrestricted matrix variables - #3510

Moreau now exposes a psd_triangle XConeSpec, so identity-pattern
SvecPSD constraints (e.g. cp.Variable((k,k), PSD=True) with no further
expression on top) can ride the x_cone path alongside nonneg / SOC.

The materialised A block for an SvecPSD constraint isn't pure
identity — entries take +1 on diagonal positions and +sqrt(2) on
off-diagonals (CVXPY's upper-triangle column-major svec convention,
matched to Moreau's svec_to_mat).  ExtractIdentityCones now
pre-populates an "expected value" per row of each candidate block
(1 for nonneg/soc/psd-diagonal, sqrt(2) for psd-off-diagonal) and
checks the materialised value against that, so the same vectorised
validity pass works for all three kinds.

MOREAU advertises:
  - SUPPORTED_CONSTRAINTS += SvecPSD
  - PSD_TRIANGLE_KIND = TriangleKind.UPPER, PSD_SQRT2_SCALING = True
  - x_cone_kinds() += 'psd_triangle'

Multi-cone SvecPSD (num_cones > 1) is left for slack-side handling
for now — it would need one XConeSpec per cone with a shared
constraint id.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

dims_to_solver_cones used to raise "Moreau does not support PSD cones"
the moment cone_dims.psd was non-empty.  That fires for any SvecPSD
constraint that ExtractIdentityCones can't take (multi-cone, e.g.
Variable((b, k, k), PSD=True), or non-identity A-block) since
ConeDims aggregates both PSD and SvecPSD constraints into psd_dims.

Moreau accepts psd_dims=[k_1, k_2, ...] on its Cones type for slack-
side PSD cones, so the right thing is to pass the dims through and let
the solver handle them — extraction still pulls the identity-pattern
single-cone case onto a psd_triangle x_cone.

Adds test_moreau_multicone_psd_via_slacks to lock in the slack-side
path against CLARABEL.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Drops the bogus num_cones() == 1 guard on SvecPSD extraction, and fixes
a latent multi-cone SOC bug where the reduction was emitting one
XConeSpec covering the entire constraint instead of one per sub-cone
(Moreau would have interpreted that as a single big SOC).

Each candidate constraint now carries its per-sub-cone row sizes (and
psd_k for SvecPSD).  When the structural-pattern check passes for the
full constraint, the flat list of x indices is split into the
sub-cone chunks and one extracted tuple is emitted per cone, all under
the same constraint id.  Moreau's solve_via_data accumulates the
per-cone KKT residual slices in iteration order and concatenates them
to produce the per-constraint dual.

The expected-per-row identity value (1 on diagonal, sqrt(2) off) is
now tiled across the num_cones sub-cones of an SvecPSD so the global
validity pass works on multi-cone constraints unchanged.

Also raises explicitly on raw PSD constraints reaching the reduction
— expand_cones is supposed to convert them to SvecPSD upstream, so
seeing one is a chain-level contract violation.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Generalises per-sub-cone metadata to a kwargs dict so extras like
alpha (PowCone3D) and alphas+dim2 (PowConeND) ride alongside psd_k
into Moreau's XConeSpec.  Formats SOC/Exp/Pow rows into the
cone-interleaved layout before scanning.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…y-cones

## Description

Stacked on #3499. This follow-up can be deferred independently of the parent PR.

For `X = cp.Variable((n, n))`, `X >> 0` constrains the symmetric part of X. This PR recognizes the two-entry off-diagonal rows produced by symmetrization and extracts their PSD cone directly. A sparse change of coordinates separates the symmetric and skew components, transforms the objective and remaining constraints, and restores the full nonsymmetric primal solution and PSD dual. Batched matrices and DPP parameter updates are covered.

Extraction reserves both entries of each coordinate pair. Parameter-dependent, overlapping, and nonmatching blocks remain on the slack side. Pair extraction also stays disabled for native coordinate bounds or integrality; Moreau's bounds are already lowered to ordinary constraints. Other scaled PSD expressions are outside this follow-up's scope.

Validation: 132 tests and 23 subtests pass across the extraction, parameter-program, DPP, Moreau, and CLARABEL suites using the local Moreau 0.4.0-beta.1 CPU build. Tests cover a known nonsymmetric optimum and PSD dual, batched parameter updates with active bounds and skew constraints, and structural rejection. Pre-commit and type-checking the changed production files pass. Full-repository pyright reports an existing optional-iterable error in `geo_mean.py`. CI inherits the parent's unavailable `moreau>=0.4.0` release dependency.

## Type of change
- [x] New feature (backwards compatible)
- [ ] New feature (breaking API changes)
- [ ] Bug fix
- [ ] Other (Documentation, CI, ...)

## [Contribution checklist](https://www.cvxpy.org/contributing/index.html#contribution-checklist)
- [x] Add our license to new files. (No new files; existing headers retained.)
- [x] Check that your code adheres to our coding style.
- [x] Write unittests.
- [x] Run the unittests and check that they are passing.
- [ ] Run the benchmarks to make sure your change does not introduce a regression.