Revert "Update vendored llama.cpp to b7847" (#14061)

llama/llama.cpp/tools/mtmd/clip-graph.h

@@ -32,6 +32,10 @@ struct clip_graph {
     const float kq_scale;
     const clip_flash_attn_type flash_attn_type;
 
+    // for debugging
+    const bool debug_graph;
+    std::vector<ggml_tensor *> & debug_print_tensors;
+
     ggml_context_ptr ctx0_ptr;
     ggml_context * ctx0;
     ggml_cgraph * gf;