!13987 [MS][LITE]fix a bug of the fc op in the fp16 subgraph and add models to the entrance guard

From: @probiotics_53
Reviewed-by: @zhanghaibo5,@zhang_xue_tong
Signed-off-by: @zhanghaibo5

mindspore/lite/src/lite_kernel.cc

@@ -105,6 +105,9 @@ int LiteKernel::PreProcess() {
 
   for (auto *output : this->out_tensors()) {
     MS_ASSERT(output != nullptr);
+    if (desc_.data_type == kNumberTypeFloat16 && output->data_type() == kNumberTypeFloat32) {
+      output->set_data_type(kNumberTypeFloat16);
+    }
     if (output->ElementsNum() >= MAX_MALLOC_SIZE / static_cast<int>(sizeof(int64_t))) {
       MS_LOG(ERROR) << "The size of output tensor is too big";
       return RET_ERROR;

mindspore/lite/test/models_tflite_fp16.cfg

@@ -149,3 +149,11 @@ scan_hms_detect_pb2tflite.tflite 1.5
 ml_location.tflite 0.5
 ml_face_openclose_tflite.tflite 0.5
 ml_object_detect_pb2tflite.tflite 1.5
+# lite-model_on_device_vision_classifier_landmarks_classifier* models' bias are caused by error accumulation and small
+# output value
+lite-model_on_device_vision_classifier_landmarks_classifier_africa_V1_1.tflite 10
+lite-model_on_device_vision_classifier_landmarks_classifier_north_america_V1_1.tflite 18
+lite-model_on_device_vision_classifier_landmarks_classifier_asia_V1_1.tflite 25
+lite-model_on_device_vision_classifier_landmarks_classifier_oceania_antarctica_V1_1.tflite 10
+lite-model_on_device_vision_classifier_landmarks_classifier_europe_V1_1.tflite 32
+lite-model_on_device_vision_classifier_landmarks_classifier_south_america_V1_1.tflite 14