Fix(gpu): add GPU fixes, dav_subspace optimization, and paged GPU memory mode

- Fix stale d_eigenvalue on GPU refresh in Diago_DavSubspace
- Fix PAGED_GPU to_range() type-punning
- Fix PAGED_GPU cal_grad reference corruption
- Fix PAGED_GPU buffer overflow and invalid D2H copy
- Fix psi_norm abort on non-resident k-points in PAGED_GPU mode
- Fix Psi PAGED_GPU issues in psi_prepare, operator(), force, and stress
- Add load_k_to_gpu before k-point loops accessing psi in force/stress/onsite
- Fix GPU memory MPI communication bugs
- Fix GPU build compilation errors (missing unique_cuda_ptr, isk parameter)
- Add dav_subspace host-buffer optimization (hcc_h/scc_h persistent buffers)
- New PAGED_GPU memory mode: device_memory_mode=input (full_gpu/paged)
- Add PsiStorageMode and psi_paging.cpp for per-k-point GPU paging
- Integrate k-point paging into HSolverPW::solve loop
- Add GPU memory mode logs in setup_psi_pw
- Add CUDA cusolver stubs and device check helpers
- Includes DFT+U PW basis fixes (issue deepmodeling#7726)

Author: dyzheng