[CausalLM] Avoid Race condition on evict experts

Avoid Race Condition on eviction experts

**Self evaluation:**
1. Build test:	 [X]Passed [ ]Failed [ ]Skipped
2. Run test:	 [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Donghak PARK <[email protected]>

Author: DonghakPark

Applications/CausalLM/layers/ernie_moe_layer.cpp

@@ -458,10 +458,14 @@ void ErnieMoELayer::incremental_forwarding(nntrainer::RunLayerContext &context,
 
 // Evict experts
 #pragma omp parallel
-    while (loaded_expert_deque.size() > 16) {
+    while (true) {
       int target_idx;
       {
         std::lock_guard<std::mutex> lock(cache_mutex);
+
+        if (loaded_expert_deque.size() > 16)
+          break;
+
         target_idx = loaded_expert_deque.front();
         loaded_expert_deque.pop_front();
         iteration_map.erase(target_idx);

Applications/CausalLM/layers/gpt_oss_moe_layer_cached.cpp

@@ -378,10 +378,13 @@ void CachedSlimGptOssMoELayer::incremental_forwarding(
 
 // Evict experts
 #pragma omp parallel
-    while (loaded_expert_deque.size() > 16) {
+    while (true) {
       int target_idx;
       {
         std::lock_guard<std::mutex> lock(cache_mutex);
+        if (loaded_expert_deque.size() > 16)
+          break;
+
         target_idx = loaded_expert_deque.front();
         loaded_expert_deque.pop_front();
         iteration_map.erase(target_idx);

Applications/CausalLM/layers/qwen_moe_layer_cached.cpp

@@ -438,10 +438,11 @@ void CachedSlimMoELayer::incremental_forwarding(
 
 // Evict experts
 #pragma omp parallel
-    while (loaded_expert_deque.size() > 32) {
+    while (true) {
       int target_idx;
       {
         std::lock_guard<std::mutex> lock(cache_mutex);
+        if (loaded_expert_deque.size() > 16) break;
         target_idx = loaded_expert_deque.front();
         loaded_expert_deque.pop_front();
         iteration_map.erase(target_idx);