llama-graph: replace ggml_repeat_4d with ggml_repeat + ggml_new_tensor_4d

The ggml_repeat_4d function doesn't exist in the current ggml API.
Replace it with the correct approach using ggml_new_tensor_4d to
create a target tensor with the desired shape, then use ggml_repeat
to repeat the input tensor to match that shape.

This fixes compilation errors when building against system-provided
ggml libraries that don't include this non-existent function.

Author: mastered-lore

src/llama-graph.cpp

@@ -770,7 +770,8 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
     if (weight_before_ffn) {
         // repeat cur to [n_embd, n_expert_used, n_tokens]
-        ggml_tensor * repeated = ggml_repeat_4d(ctx0, cur, n_embd, n_expert_used, n_tokens, 1);
+        ggml_tensor * target_shape = ggml_new_tensor_4d(ctx0, cur->type, n_embd, n_expert_used, n_tokens, 1);
+        ggml_tensor * repeated = ggml_repeat(ctx0, cur, target_shape);
         cur = ggml_mul(ctx0, repeated, weights);
         cb(cur, "ffn_moe_weighted", il);
     }