Request sufficient accuracy in three gradient checks

[SteveDiamond]

Description

Three finite-difference gradient checks fail with SCS 3.3.1 at default solve accuracy: inequality duals, vector-equality duals, and the SDP example. Their finite-difference step is smaller than the solver error, so perturbations can produce a numerical Jacobian that disagrees with the analytical derivative.

Request eps=1e-10 at those three call sites, following the existing gradient tests that already specify this tolerance. The gradient assertions and finite-difference settings are unchanged. This is six added lines and three removed lines across two test files; there are no dependency caps or production-code changes.

Found while investigating CVXPY CI in cvxpy/cvxpy#3513. CVXPY's extension workflow clones this repository, so its remaining three failures need this upstream change.

Validation

• Reproduced all three failures locally with SCS 3.3.1 before the change; all three pass afterward.
• pytest tests/test_dual_variables.py tests/test_torch.py: 50 passed, 28 skipped (optional backends), against CVXPY master plus Fix SCS 3.3 compatibility and CI accuracy assumptions cvxpy#3513.
• Ruff passes on both changed files.
• Kept as a draft for upstream CI and review.