Add audio data provider and preprocessor for speech recognition datasets.

[xinghai-sun]

• Prepare one or more public English speech recognition data sets (e.g. LibriSpeech), and respective baselines.
• Convert all audio file formats to .wav format.
• Add a file manifest generator for each dataset, and add a merger if there exist more than one datasets. Make this interface unified across different datasets.
• Add spectrogram feature extractor, power normalizer etc.
• Add transcription text parser (tokenization, dictionary generation etc).
• Add batch data reader with SortaGrad.
• Refer to the DS2 design doc and update it when necessary.
• Please pull your codes and docs into PaddlePaddle/models.

[xinghai-sun]

Pull Request PaddlePaddle/models#55