Fix DLRM profiler argument passing

[kiwigitops]

Summary

• pass the profiler device selection by keyword because profiler options after enabled are keyword-only
• select cuda when the DLRM run is using a GPU and None for CPU runs
• preserve the existing disabled-profiling behavior while avoiding the startup TypeError

Fixes #2589

Validation

• Reproduced the original profile.__init__() positional-argument failure on current main with PyTorch 2.13.0
• Ran the issue's random-data CPU path after the change and completed all three training iterations
• Repeated with --enable-profiling; the run completed and produced nonempty profiler table and Chrome trace files
• Parsed the Chrome trace and verified traceEvents, schemaVersion, and deviceProperties
• python3 -m py_compile torchbenchmark/models/dlrm/dlrm_s_pytorch.py
• git diff --check