Another bug in Star Transformer

RT, another bug

fix mask bug in star-transformer

fix the bug described in #138 . Thank @wlhgtc for bug reporting and pr.

If you use masked_fill according to ex_mask (0 for pad), it will fill not padding position(which value in ex_mask is 1) with 0, this will lead a bad performance.

* 修复fast_param_mapping的一个bug
* Trainer添加自动创建save目录
* Vocabulary的打印，显示内容

* 给callback添加给定几个只读属性
* 通过manager设置这些属性
* 代码优化，减轻@transfer的负担

Add Star-Transformer

* add ignore_type in DataSet.apply

fix tutorial link

[new] Add ENAS (Efficient Neural Architecture Search)

fastNLP V0.3.1

* 升级parser API和模型
* update docs: add new pages for tutorials
* upgrade CWS api download source
* add a new method for dataset field access
* add introduction for bert
* add more unit tests for api/processor
* remove unused test data. Add new test data.

* load pre-trained BERT weights from local binary
* add tests

* rename callback methods. Use fastai's notation.
* add a new callback method - on_valid_begin