Refactor CUDA backend: replace CudaInternal singleton with HostSharedPtr default_instance

[Copilot]

Refactors CudaInternal to match the HIP backend pattern established in kokkos#7646. Replaces the static singleton with HostSharedPtr-managed lifetime, moves init/finalize logic into constructor/destructor.

Changes

• CudaInternal class (Kokkos_Cuda_Instance.hpp):

  • static CudaInternal& singleton() → static HostSharedPtr<CudaInternal> default_instance
  • Default constructor → CudaInternal(cudaStream_t stream) (initialization in ctor)
  • finalize() cleanup logic → ~CudaInternal() (with fence before destruction)
  • Removed initialize(), finalize(), is_initialized(), was_initialized, was_finalized
  • Copy/assign marked = delete

• Cuda execution space (Kokkos_Cuda_Instance.cpp):

  • impl_initialize(): creates default_instance via HostSharedPtr(new CudaInternal(stream), deleter)
  • impl_finalize(): default_instance = nullptr
  • Cuda(): copies default_instance (was: non-owning ptr to singleton)
  • Cuda(stream, manage_stream): new CudaInternal(stream) with conditional stream-destroying deleter

• Kokkos_CudaSpace.cpp: CudaInternal::singleton(). → CudaInternal::default_instance->

Pattern (mirrors HIP)

// init
Impl::CudaInternal::default_instance = Impl::HostSharedPtr(
    new Impl::CudaInternal(stream), [](Impl::CudaInternal *ptr) {
      cudaStream_t s = ptr->m_stream;
      delete ptr;
      KOKKOS_IMPL_CUDA_SAFE_CALL(cudaStreamDestroy(s));
    });

// finalize
Impl::CudaInternal::default_instance = nullptr;

// default Cuda instance copies the shared ptr
Cuda::Cuda() : m_space_instance(Impl::CudaInternal::default_instance) {}

Original prompt

I want the CUDA backend refactored so that the CudaInternal class doesn't have the singleton member, and that the initialization and finalization happens in constructor and destructor respectively. Instead of the singleton it should use a static HostSharedPtr member in the CudaInternal class to track the default instance. This refactor should follow the refactor done for the HIP backend in the PR 7646 (kokkos#7646).

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.