Fix device parametrization for optimizer benchmarks

[emme1t]

Fixes #2084.

test_benchmark_blockwise accepts device without parametrizing it, so selecting optimizer benchmarks fails during setup with 27 fixture 'device' not found errors. Parametrize it with the existing get_available_devices() helper and use the same CPU/Windows paged-optimizer skips as test_optimizer32bit.

Validation on Windows, Python 3.12.13 and PyTorch 2.14.0+cpu:

• python -m pytest tests/test_optim.py -k test_benchmark_blockwise -m benchmark --setup-only -q: reproduced 27 setup errors on main; all 27 cases set up successfully with the patch.
• Executed the existing benchmark functions with 64x64 tensors, one CPU thread and TORCHDYNAMO_DISABLE=1, retaining the 500-step loops: 12 passed, 15 skipped. This exercises the 8-bit state path at 4096 elements. Tensor sizes were adjusted by a local pytest collection plugin; the committed benchmark dimensions remain 4096x4096.
• The same setup-only command with BNB_TEST_DEVICE=cuda successfully resolves the explicit device selection.
• Changed-file pre-commit hooks and git diff --check pass.

FreeBSD and accelerator execution were not available locally. Full-size benchmark timings were not measured.