* update to paddle 2.

* update readme

* upgrade multi card fintune example

* use paddle.AdamW, use grad acc

* bump propeller

* remove grad acc

* fix ner

* update propeller & distribued sample

* wip

* +seq2seq

* format

* fix erneigen

* fix pretrain

* fix static

* update propeller  for py37 compat

* fix pretrain static

* up readme

* update readme

* static pretrain

* remove optimization out of core libray

* ner use `cross_entropy`, use `ignore index`

* fix dygraph pretrain: add stop criteria

* bugfix, LN wrong initialize

* add grad acc for classifiction task

* seq2seq use fp32 when decoding

* use `paddle.io.DataLoader`

* + distill

* update readme

* update distill fig link

* propeller use vdl

* do not use pure fp16 for static graph
Co-authored-by: Nchenxuyi <work@yq01-qianmo-com-255-129-11.yq01.baidu.com>

1. finetune/ pretrian fp16 amp
2. finetune/ pretrain multi processing
3. ner infer
4. py3 compat

ERNIE 2.0 is a continual pre-training framework for language understanding in which pre-training tasks can be incrementally built and learned through multi-task learning