Matchzoo-py

- add automation modules
- mz.auto.tuner that automatically search for model hyper parameters
- mz.auto.preprer that unifies model preprocessing and training processes
- add QuoraQP dataset
- rewrite mz.DataGenerator to be callback-based
- fix models behaviors under classification tasks
- reorganize project structure, the most significant one being moving `processor_units` to `preprocessors.units`
- rename redundant names (e.g. NaiveModel -> Naive, TokenizeUnit -> Tokenize)
- update the tutorials
- various other updates

- Add new models: MVLSTM, MatchPyramid, HBMP, DUET, MatchSRNN, aNMM and DIIN
- Add activation to classification models
- Add Ngram callbacks for `Dataset`
- Remove `DSSMPreprocessor`, `CDSSMPreprocessor` and `CDSSMPadding`
- Add BERT `attention_mask` and `token_type_ids`
- Support specifying GPUs for DataParallel and remove parameter `data_parallel` in `trainer`
- Fix other bugs

- Written based on PyTorch
- Add BertPreprocessor, BertPadding, BertModule and Bert model
- Speed up loading embedding & support word2vec, GloVe and fastText
- Implement a majority of matching models
- Support dynamic padding
- Auto check task when defining a task
- Support multiprocess DataLoader
- Support early stopping, clipping gradient norm, validation interval and auto save best model