Add CUDA backend

[gitctrlx]

Add CUDA Backend

Implement a CUDA backend for the candy framework to enable high-performance computation leveraging NVIDIA GPUs. This backend should support tensor operations, neural network layers, and necessary data movement across host and device. It should be designed for extensibility to facilitate future support of more CUDA features and APIs.

Tasks

• Define API for integrating CUDA backend into candy
• Implement required tensor operations (allocation, data transfer, core arithmetic, etc.)
• Support basic neural network layers and operator primitives using CUDA
• Ensure compatibility with candy's existing backend management
• Introduce error handling, resource management, and clear documentation
• Add tests and benchmarks

Reference: Review CUDA integration in other ML frameworks and candy's existing backend design for alignment and inspiration.