feat: add CUDA graph capture modes

[honeyspoon]

Why

cuda-async always begins stream capture in relaxed mode. Callers that need CUDA to detect interfering work from the current thread or process currently have to bypass the graph API.

What changed

• Add CaptureMode::{Global, ThreadLocal, Relaxed}.
• Add explicit-mode variants for operation and scoped graph capture.
• Re-export CaptureMode from the prelude.
• Keep existing capture methods unchanged, with relaxed mode as their default.
• Exercise both capture paths with every mode in the GPU-gated graph tests.

This does not change graph instantiation, upload, replay, or ownership behavior.

Validation

• cargo fmt --all -- --check
• cargo check -p cuda-async --all-targets
• cargo test -p cuda-async --test cuda_graph --no-run
• cargo test -p cuda-async --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.