Fix IO benchmark storage setup and defer GPU detection

[SoraSuegami]

Summary

• Reinitialize storage against the active bench directory before building the naive vector encoding benchmark for both diamond_io and aky24_io.
• Defer GPU synchronization and device detection until after error simulation selection, so search-only runs can proceed without requiring a local GPU.
• Preserve the selected CRT-depth-driven benchmark flow while keeping the H200 benchmark path aligned with the latest local fix.

Testing

• cargo +nightly fmt --all
• FIDESLIB_SKIP_SUBMODULE_UPDATE=1 cargo test --release --features gpu --test test_gpu_diamond_io --test test_gpu_aky24_io --no-run
• Verified the affected H200 benchmark rerun completed past the previous public-LUT checkpoint failure