CUDA: Simplify and improve CUDA graphs through use of indirect copy pointers

Previously there was complexity in the CUDA graphs implementation due
frequently changing parameters to copy kernels associated with K and V
cache pointers. This patch simplifies by using indirection to avoid
such parameters frequently changing, avoiding the need for frequent
graph updates.

Fixes #12152

Author: agray3

ggml/include/ggml-backend.h

@@ -349,6 +349,9 @@ extern "C" {
     GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
+    // Copy destination pointers for copy operations pointers to backend
+    GGML_API void ggml_backend_dest_ptrs_copy(char ** host_dest_ptrs, const int host_dest_ptrs_size);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/src/ggml-cuda/common.cuh

@@ -711,7 +711,7 @@ struct ggml_cuda_graph {
     bool disable_due_to_failed_graph_capture = false;
     int number_consecutive_updates = 0;
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
-    std::vector<char **> updated_kernel_arg;
+    std::vector<char *> cpy_dest_ptrs;
 #endif
 };