ggml CUDA Performance and Synchronization Optimization

Through the b9820 update to the ggml library, CUDA backend performance has been improved.

The key changes are as follows:

• Reduced synchronization overhead: Optimized performance by reducing the number of synchronizations during the split compute process and relaxing inter-token synchronization requirements.
• Asynchronous copy support: CPU-to-CUDA asynchronous copying is now possible via the ggml_backend_cuda_cpy_tensor_async() function, improving data transfer efficiency.
• Backend structure improvements: Rewrote the backend detection logic to prevent link conflicts, and introduced a synchronization relaxation structure that can also be utilized by other backends such as Vulkan.
• Code stability: Resolved link issues by relaxing the check logic for asynchronous CUDA copies, and carried out overall code cleanup.