mtmd: add clip_graph::build_mm() (#20751)

* clip: add build_mm()

* apply to all models

* add TODO for bias overload

tools/mtmd/clip-graph.h

@@ -41,6 +41,11 @@ struct clip_graph {
     virtual ~clip_graph() = default;
     virtual ggml_cgraph * build() = 0;
 
+    // wrapper around ggml_mul_mat, allow hooking (e.g. LoRA, clamping) depending on the model
+    // tensor w should be the weight matrix, and tensor x should be the input
+    virtual ggml_tensor * build_mm(ggml_tensor * w, ggml_tensor * x) const;
+    // TODO: build_mm(w, b, x) to support bias
+
     //
     // utility functions
     //