Fix: route GEMMA4 output_norm to ctx_output instead of ctx_output_split

I mistakenly added routed
output_norm for GEMMA4 and GEMMA4_MTP to ctx_output_split.
output_norm is a 1D vector that gains no benefit from being split
across GPUs and incurs unnecessary communication overhead.
Reverted it to ctx_output instead.

Author: Nexesenex

src/llama-load-tensors.cpp

@@ -2153,7 +2153,7 @@ bool create_tensors_helper::create_gemma4_tensors(const LLM_TN & tn) {
         model.per_layer_proj_norm  = create_tensor(ctx_output, tn(LLM_TENSOR_PER_LAYER_PROJ_NORM,  "weight"), {n_embd_per_layer}, 0);
     }
 
-    model.output_norm = create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
 
     int rope_freqs_flag = 0;
 
@@ -2239,7 +2239,7 @@ bool create_tensors_helper::create_gemma4_mtp_tensors(const LLM_TN & tn) {
     }
 
     model.tok_embd    = create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
-    model.output_norm = create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
     model.output      = create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
     auto output_extra = create_tensor(ctx_output_split, "output_extra.weight", {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
     if (output_extra) {