Avoid cuBLAS in L2 persistence benchmark

[msaroufim]

Summary

Generate the L2 persistence test's expected output with an elementwise reduction instead of the benchmark helper that dispatches through cuBLAS. The benchmarked kernels still use the single-column GEMM path from #51; only expected-output setup avoids cuBLAS.

Root cause

The main-branch CI failure after #51 happened at expected-output generation, before timing started:

expected = _gemm_vec(w, x)
torch.mm(w, x_col)
RuntimeError: CUDA error: CUBLAS_STATUS_INVALID_VALUE when calling cublasSgemm(...)

That showed the instability was using cuBLAS during test-case setup, not specifically GEMV.

Validation before opening

• python3 -m py_compile test/l2_persistence.py
• git diff --check
• Full Modal L4 suite using the failed 07424cc/run 27250846501 wheel artifact passed, including l2_persistence.py with OK: L2 persistence confers no benchmark advantage.

I am leaving this as draft until the PR's own GitHub Actions run is green.