llama.cpp b9890 refactors cuBLAS and fixes CDNA + BF16 logic

The llama.cpp project released version b9890, which includes significant backend refactoring and bug fixes. The update removes the -sm row for CUDA and refactors cuBLAS integration to improve stability.

• Fixes CDNA + BF16 logic errors
• Corrects bad return values in operations
• Resolves issues with src0 strides and contiguous memory requirements
• Fixes GGML_CUDA_FORCE_CUBLAS configuration
• Corrects casts to BF16 format

This release provides updated binaries for macOS, Linux, Windows, Android, and openEuler across CPU, GPU, and various accelerator backends.