Harden QR checker and benchmark cases

[msaroufim]

Summary

• Add heterogeneous mixed QR inputs that interleave well-conditioned and ill-conditioned matrices inside one batch.
• Add mixed and ill-conditioned benchmark cases so robustness affects ranking, not just correctness gating.
• Check the LAPACK-style factor residual per matrix, so one large/easy matrix cannot hide a failed matrix in the same batch.
• Explicitly reject NaN/Inf materialized Q, R, Q.T @ A, Q.T @ Q, and Q @ R paths before residual comparisons.

Validation

• python3 -m py_compile problems/linalg/qr_py/reference.py problems/linalg/qr_py/eval.py problems/linalg/qr_py/submission.py problems/linalg/qr_py/task.py
• git diff --check

Notes

• I did not run any production admin problem update command.
• I found 6 existing QR submissions in the production DB with public passing B200 scores and failed secret B200 runs; this PR is meant to make that class of hidden-case failure visible in the problem definition.