cuda.core: support clustered and cooperative kernel graph nodes

[Andy-Jost]

Background

cuda.core.launch() supports LaunchConfig.cluster and LaunchConfig.is_cooperative, but explicit GraphDefinition kernel-node construction currently copies only the basic grid, block, and shared-memory fields. Cluster and cooperative attributes are ignored, and a clustered grid is incorrectly treated as a block grid. KernelNode.update() likewise does not model the corresponding kernel-node attributes or clustered-grid semantics.

This is a follow-up to #2352 and #2395.

Scope

Add end-to-end support for clustered and cooperative graph kernel nodes, including:

• Explicit construction using clustered and cooperative LaunchConfig values.
• Correct conversion between logical cluster-grid dimensions and CUDA block-grid dimensions.
• Setting and preserving the corresponding CUDA kernel-node attributes.
• Reconstruction and inspection of explicit and captured nodes.
• Partial updates of launch configuration, kernel, and arguments without losing existing attributes.
• Device and kernel capability validation with clear errors before graph mutation.
• Public API documentation and release notes.

Acceptance criteria

• Clustered and cooperative nodes can be explicitly constructed, instantiated, and launched successfully on supported devices.
• Captured nodes reconstruct with their launch dimensions and kernel-node attributes intact.
• KernelNode.config faithfully reports clustered and cooperative state.
• Partial updates preserve omitted dimensions, attributes, kernel arguments, and ownership metadata.
• Unsupported devices, kernels, or attribute combinations fail before mutating the graph.
• Tests cover explicit and captured graphs, clustered-grid conversion, cooperative launch, reconstruction, partial updates, capability validation, and failed-update atomicity.