Fix performance of WeightedRandomSampler (pytorch#10636)

Summary:
Since pytorch#8958 was merged, the BatchSampler samples 0d tensors from WeightedRandomSampler instead of integers. It significantly reduces performance. This PR fix it the same way as pytorch#10361 fix DistributedSampler.
Pull Request resolved: pytorch#10636

Differential Revision: D9423869

Pulled By: zou3519

fbshipit-source-id: f94da2d4cccf70e63beea6cfc3d1230b5610ae44

Author: Chetter2

torch/utils/data/sampler.py

@@ -95,7 +95,7 @@ def __init__(self, weights, num_samples, replacement=True):
         self.replacement = replacement
 
     def __iter__(self):
-        return iter(torch.multinomial(self.weights, self.num_samples, self.replacement))
+        return iter(torch.multinomial(self.weights, self.num_samples, self.replacement).tolist())
 
     def __len__(self):
         return self.num_samples