A couple tweaks to ModelLoaderRegistry to improve potential contention.

1) Synchronize only the inner method that accesses the cache, as the outer call
that loops and filters the loaders appears to not have threading issues.
2) Probably less of an impact since I imagine the loops are generally small,
but only allocate an ArrayList once you find a matching loader, and allocate it
with size - i which is the max possible remaining in the loop, rather than
always allocating it to size.

-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=205877123

library/src/main/java/com/bumptech/glide/load/model/ModelLoaderRegistry.java

@@ -67,14 +67,19 @@ public class ModelLoaderRegistry {
   }
 
   @NonNull
-  public synchronized <A> List<ModelLoader<A, ?>> getModelLoaders(@NonNull A model) {
+  public <A> List<ModelLoader<A, ?>> getModelLoaders(@NonNull A model) {
     List<ModelLoader<A, ?>> modelLoaders = getModelLoadersForClass(getClass(model));
     int size = modelLoaders.size();
-    List<ModelLoader<A, ?>> filteredLoaders = new ArrayList<>(size);
+    boolean isEmpty = true;
+    List<ModelLoader<A, ?>> filteredLoaders = Collections.emptyList();
     //noinspection ForLoopReplaceableByForEach to improve perf
     for (int i = 0; i < size; i++) {
       ModelLoader<A, ?> loader = modelLoaders.get(i);
       if (loader.handles(model)) {
+        if (isEmpty) {
+          filteredLoaders = new ArrayList<>(size - i);
+          isEmpty = false;
+        }
         filteredLoaders.add(loader);
       }
     }
@@ -92,7 +97,8 @@ public class ModelLoaderRegistry {
   }
 
   @NonNull
-  private <A> List<ModelLoader<A, ?>> getModelLoadersForClass(@NonNull Class<A> modelClass) {
+  private synchronized <A> List<ModelLoader<A, ?>> getModelLoadersForClass(
+      @NonNull Class<A> modelClass) {
     List<ModelLoader<A, ?>> loaders = cache.get(modelClass);
     if (loaders == null) {
       loaders = Collections.unmodifiableList(multiModelLoaderFactory.build(modelClass));