Fix prewarming (#4454)

Summary:
Prewarms the model if `config.should_prewarm_model` is `true`. This improves the latency of first inference call as the necessary objects are created when the model is prewarmed.

Testing:
Existing tests

Pull Request resolved: #4454

Reviewed By: kirklandsign

Differential Revision: D60469148

Pulled By: cccclai

fbshipit-source-id: d88883e721269d03298265dd420f08cbbe4787ce

Author: cymbalrush

backends/apple/coreml/runtime/delegate/ETCoreMLModelManager.mm

@@ -655,7 +655,7 @@ - (void)prewarmRecentlyUsedAssetsWithMaxCount:(NSUInteger)maxCount {
 
             NSError *prewarmError = nil;
             if (![asset prewarmAndReturnError:&prewarmError]) {
-                ETCoreMLLogError(localError,
+                ETCoreMLLogError(prewarmError,
                                  "%@: Failed to prewarm asset with identifier = %@",
                                  NSStringFromClass(strongSelf.assetManager.class),
                                  asset.identifier);

backends/apple/coreml/runtime/delegate/backend_delegate.mm

@@ -157,7 +157,7 @@ - (BOOL)_loadAndReturnError:(NSError * _Nullable __autoreleasing *)error {
     if (self.config.should_prewarm_asset) {
         [modelManager prewarmRecentlyUsedAssetsWithMaxCount:1];
     }
-    
+
     return YES;
 }
 
@@ -188,9 +188,14 @@ - (ModelHandle*)loadModelFromAOTData:(NSData*)data
         return nil;
     }
 
-    return [self.impl loadModelFromAOTData:data
-                             configuration:configuration
-                                     error:error];
+    auto handle = [self.impl loadModelFromAOTData:data
+                                    configuration:configuration
+                                            error:error];
+    if ((handle != NULL) && self.config.should_prewarm_model) {
+        [self.impl prewarmModelWithHandle:handle error:nil];
+    }
+
+    return handle;
 }
 
 - (BOOL)executeModelWithHandle:(ModelHandle*)handle