Use GEMM in L2 persistence CI test

[msaroufim]

Summary

Update the L2 persistence regression test to use a single-column GEMM (torch.mm(w, x[:, None])) instead of GEMV (torch.mv). The RHS remains logically vector-shaped for the benchmark, but keeping it rank-2 avoids the flaky cublasSgemv path that failed in Modal L4 CI.

Root cause

The failing GitHub Actions job hit CUBLAS_STATUS_INVALID_VALUE inside torch.mv(w, x) while generating expected outputs for test/l2_persistence.py. The previous lazy CUDA initialization fix moved the failure site but did not remove the fragile cuBLAS GEMV dispatch.

Validation

• python3 -m py_compile test/l2_persistence.py
• git diff --check
• Full Modal L4 GPU test suite using the failed workflow wheel artifact from run 27183274114: passed, with l2_persistence.py reporting hacked/honest median ratio = 0.998 and OK: L2 persistence confers no benchmark advantage.
• Narrowed Modal L4 l2_persistence.py rerun 3 additional times with the same failed workflow wheel artifact: all passed.

Notes

T4 is not a useful target for this regression test because Turing does not support the persisting-L2 behavior being exercised. L4 remains the intended CI target for this check.