fix(cuda): stabilize AL-IPC and harden release compatibility

[KemengHuang]

Summary

• harden the current AL-IPC active-set integration with type-correct earliest-TOI filtering, decay-derived pair lifetime, persistent amortized CUDA scratch, and AL-specific CCD handling
• correct sample-88 trajectory failures by honoring newton/semi_implicit/K_min, fixing the EE friction Jacobian, rejecting and rolling back failed line-search trials, and restoring stable mass-based per_vertex penalty scaling as the default
• preserve the owner-required AL parallel-EE policy: the negative mollifier coefficient keeps need_mollify() false
• harden the merged fix(cmake): quote CUDA architecture list so wheels ship every target #490 architecture fix with one checked CMake setter covering the final backend, every component OBJECT target, and the CUDA test target
• distinguish packaged SASS from PTX-JIT compatibility in uipc doctor; CUDA 12.8.1 PTX fallback now uses Linux 570.124.06 / Windows 572.61 driver floors while native SASS retains the CUDA 12.x minimum-driver floor
• update public configuration/build documentation, ADRs, handoff notes, release policy, and wheel smoke checks

Behavior changes

• AL-IPC consumes newton/semi_implicit/enable and K_min for cumulative safe-path progress; beta_tol remains IPC-only
• contact/al-ipc/mu_scale_mode defaults to per_vertex; diag_norm remains available as an experimental explicit mode
• an exhausted AL line search restores the recorded start point instead of accepting an energy-increasing trial
• compatibility diagnostics report whether a GPU selects SASS or PTX and apply the corresponding driver requirement; this is diagnostic only and does not add a runtime driver gate

Validation

• sample 88 with cloth: 250/250 frames converged, no line-search/Newton limit or runtime error
• matched 60-frame sample-88 centroid error versus IPC reduced from 0.2963/0.0807 to 0.0115/0.0045
• CMake CUDA backend: 22 cases / 335 assertions
• Core: 36 cases / 1040 assertions
• single-process simulation suite: 95 cases / 14213 assertions
• comma-form release architecture configure: all 214 test-enabled CUDA translation units carry all six codegen entries
• repository contracts: 48/48
• portable Python: 80 passed / 1 skipped
• fast CTest: 3/3
• CMake and XMake production CUDA targets built successfully
• Python 3.14 uipc doctor --probe-cuda passed on RTX 5090 / CUDA 13.2
• complete Doxygen/MkDoxy/MkDocs build passed

The tracked libuipc-samples gitlink is unchanged; local sample edits are not part of this PR.