Add Cholesky mini-training validation contract

[msaroufim]

Summary

• opt Cholesky into application validation with a two-field contract:
  version and script
• keep the workload, shapes, thresholds, and result construction inside the
  problem-owned script
• run a deterministic natural-gradient logistic-regression loop over eight B200
  shapes
• compare convergence, factor reconstruction, and synchronized wall time
  against torch.linalg.cholesky_ex

Top-10 users are trusted here. Kernelbot supplies each user's best eligible
submission. The workload does not duplicate KernelGuard and does not reject
PyTorch or hybrid implementations merely for their execution route.

Debug evidence

The attached Seraphim kernel completed a fresh current-tip isolated B200 run:

• 5/8 shapes validated
• 2.82x geomean synchronized-wall speedup across measured shapes
• n=32, 64, 128, 256, 1024: stable, convergent, and faster than PyTorch
• n=512, 2048, 4096: rejected for non-finite factors

Ruff, Python byte-compilation, contract loading, and the end-to-end B200 run
passed.

Depends on gpu-mode/kernelbot#518, which is merged and deployed.