Fix Clarabel PSD cone permutation for multiple cones

[bstellato]

Summary

Fix two bugs in solve_internal() when using Clarabel with PSD cones.

Problem

When using diffcp.solve_and_derivative(..., solve_method='CLARABEL') with problems containing multiple PSD cones, the solution was incorrect. For example, a problem with optimal objective ~0.088 would return ~1.3e-10.

Root Cause

Two bugs in the Clarabel code path:

Bug 1: Wrong start_row increment (line 542)

# Before (incorrect)
start_row += v

# After (correct)
start_row += v * (v + 1) // 2  # triangular number for vectorized PSD cone

For a PSD cone of dimension n (an n×n matrix), the vectorized representation has n*(n+1)/2 elements, not n. This caused incorrect row permutations when problems have multiple PSD cones.

Bug 2: Missing inverse permutation for y and s

After solving with Clarabel, the returned y and s vectors are in Clarabel's upper-triangular convention but were not permuted back to SCS's lower-triangular convention.

Added inverse_permute_psd_solution() function to handle this conversion.

Testing

Added comprehensive test file tests/test_clarabel_psd.py with 6 tests:

• test_multiple_psd_cones_objective_match - verifies fix for the reported issue
• test_single_psd_cone - single PSD cone works correctly
• test_mixed_cones - problems with zero, nonneg, and PSD cones
• test_constraint_satisfaction - verifies Ax + s = b
• test_derivative_lsqr_mode - derivatives work in lsqr mode
• test_psd_permutation_logic - permutation logic is correct

Verification

Before fix:

CVXPy objective: 0.088146
diffcp + Clarabel objective: 1.3e-10  # WRONG

After fix:

CVXPy objective: 0.088146
diffcp + SCS objective: 0.088158
diffcp + Clarabel objective: 0.088146  # CORRECT