Fix L2 persistence Modal CI test

[msaroufim]

Summary

Move the L2 persistence test's CUDA device queries, libcudart loading, and large weight allocation behind lazy process-local initialization. This avoids initializing CUDA at module import before the isolated benchmark subprocess/thread runs.

Root cause

The Modal L4 CI job failed in test/l2_persistence.py when torch.mv(_W, x) raised CUBLAS_STATUS_INVALID_VALUE. The test created _W and queried CUDA globals at import time, then the benchmark harness spawned an isolated process and ran setup inside the native benchmark thread. Deferring that CUDA state creation keeps the test setup inside the process that uses it.

Validation

• python3 -m py_compile test/l2_persistence.py ci/modal_gpu_test.py python/pygpubench/__init__.py
• git diff --check
• Reran the same Modal L4 test command using the failed workflow wheel artifact; l2_persistence.py completed with hacked/honest median ratio = 1.001 and OK: L2 persistence confers no benchmark advantage.