Add NaN check to prevent false singular matrix errors

[lchen2331]

Fix issue: #2667
The issue is caused by oneMKL receiving NaN inputs, which can lead to out-of-bounds memory access and convert NaN to 0, causing the singular check to fail. A NaN check is added before oneMKL computation, and NaN outputs are returned when NaN inputs are detected to align with cuda results.

Testing:

pytest -sv test/regressions/test_linalg_solve_nan.py

Output:

test/regressions/test_linalg_solve_nan.py::TestLinalgSolveNaN::test_solve_batch_mixed_nan PASSED
test/regressions/test_linalg_solve_nan.py::TestLinalgSolveNaN::test_solve_cayley_transform PASSED
test/regressions/test_linalg_solve_nan.py::TestLinalgSolveNaN::test_solve_nan_variants PASSED

[Copilot]

Pull request overview

This PR adds NaN input validation to prevent false singular matrix errors in MKL linear algebra operations. When NaN values are detected in inputs, the functions now return NaN outputs with appropriate default values instead of passing invalid data to oneMKL, which could cause out-of-bounds memory access.

Changes:

• Add NaN checks in lu_solve_mkl and lu_factor_mkl functions
• Return NaN-filled outputs when NaN inputs are detected
• Include necessary headers for isnan and arange operations

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Performance outliers, please check!

• 🟡 [80%, 90%), may be fluctuations

Category	Model	Target vs. Baseline [Eager]	Target vs. Baseline [Inductor]
torchbench_bfloat16_training	resnet18	0.94745	0.899606

[CuiYifeng]

Please also add test cases for input with nan, thanks.