Implement a handle-based resource management layer #1303
Description
Implement a new handle-based resource management layer in the cuda.core Python package, replacing direct lifetime management via __del__, __dealloc__, and similar destructors with the handle abstraction described in the attached proposal. The task includes updating existing CUDA resource wrappers (e.g., contexts, streams, device pointers) to own and interact with opaque handles rather than raw CUDA resources, ensuring referentially transparent lifetimes, structural dependency tracking, and correct automatic release when handles become unreachable.
Proposal_ Handle Abstraction for CUDA Resource Management in cuda-core.pdf