Add KSP and classical AMG stage test coverage

[LiRayX]

Summary

• restore unit-test source discovery and CI failure propagation
• add convergence, zero-residual, and iteration-limit coverage for PCG, GMRES, and PBiCGStab
• add focused classical AMG coverage for strength, coarsening, interpolation, smoothers, scaling, and the RAP operator
• exercise host/device precision variants where supported and both AMGX/cuSPARSE RAP backends

Algorithm coverage

• Strength: AHAT, ALL, AFFINITY
• Coarsening: PMIS, HMIS, aggressive PMIS/HMIS, CR
• Interpolation: D1, D2, MULTIPASS, including a long-range multipass case
• Smoothers: block Jacobi, L1 Jacobi, Chebyshev, Chebyshev polynomial, GS, multicolor GS/DILU
• Scaling: diagonal symmetric, binormalization, nbinormalization
• RAP: comparison against an independent dense Galerkin reference

Fixes exposed by the new tests

• make device strength=ALL mark every connection consistently
• remove the unconditional process exit and implement UNSCALE in device binormalization
• correct normalization-factor folding in binormalization and nbinormalization
• keep the legacy multicolor-GS Poisson test on its supported inline-diagonal layout

Validation

• CUDA 13.3, GCC 13, CUDA architecture 120
• incremental build succeeds
• focused combined regression: 62/62 tests passed
• git diff --check passes