Implement CUDA-aware IO in ADIOS2

[ia267]

Current implementation treats CPU and GPU data identically, which may lead to unnecessary data transfers using the CUDA backend

Implementation details:

• Modify adios2_io.f90 to detect backend type at runtime
• Add backend-specific implementation for write operations
• Implement CUDA-specific code path using ADIOS2's GPU-aware functionality
• Add proper synchronisation for CUDA streams before I/O
• Add build-time detection of ADIOS2 CUDA support

Acceptance criteria:

• Data remains on GPU during I/O operations when using CUDA backend
• Performance improvement for checkpoint/snapshot operations with large datasets on GPU
• Graceful fallback when ADIOS2 lacks CUDA support
• No performance regression for OMP backend