example.tts: Update tacotron2 inference interface

There are two updates in Nvidia Tacotron2 interface:
1. infer() receives input_lengths argument
2. infer() returns mel, mel_lengths, alightments

examples/text_to_speech_synthesizer/waveglow_handler.py

@@ -91,7 +91,8 @@ def preprocess(self, data):
 
     def inference(self, data):
         with torch.no_grad():
-            _, mel, _, _ = self.tacotron2_model.infer(data)
+            input_lengths = torch.IntTensor([data.size(1)]).to(device=self.device, dtype=torch.int64)
+            mel, _, _ = self.tacotron2_model.infer(data, input_lengths)
             audio = self.waveglow_model.infer(mel)
 
             return audio