feat: expose CUDA device attributes

[honeyspoon]

Why

Device exposes its L2 size, but querying related occupancy, bandwidth, and persisting-cache limits requires direct calls through cuda_core::sys.

What changed

• Add a non-exhaustive DeviceAttribute enum.
• Add safe Device::attribute access to integer-valued CUDA device properties.
• Cover compute capability, clocks, memory bus and cache limits, and per-SM occupancy resources.
• Route l2_cache_size_bytes through the shared query path.
• Add a GPU-gated test that queries every exposed attribute and checks the existing L2 helper against the generic result.

This extends the existing Device API only; it does not add context, stream, module, or launch abstractions.

Validation

• cargo fmt --all -- --check
• cargo check -p cuda-core --all-targets
• cargo test -p cuda-core --test device_attributes --no-run
• cargo test -p cuda-core --doc
• full workspace Clippy, compile-only, and CPU suites passed before splitting the independent changes

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.