graph : normalize Q, K, V shapes + sync cross attention (#12449)

* graph : normalize Q, K, V shapes and add comments

ggml-ci

* context : synchronize before getting cross attention data

* model : fix command-r attention norm check

src/llama-context.cpp

@@ -1143,6 +1143,8 @@ int llama_context::encode(llama_batch & inp_batch) {
     if (model.arch == LLM_ARCH_T5 && t_embd) {
         //cross.t_embd = t_embd;
 
+        synchronize();
+
         cross.n_embd = t_embd->ne[0];
         cross.n_enc  = t_embd->ne[1];
         cross.v_embd.resize(cross.n_embd*cross.n_enc);