Support setting warmup start lr and eta_min in Cosine.
Fix bug that Linear can not decay to end_lr when setting warmup.

* add static training

* fix typo

* add se fp16

* rm note

* fix loader

* fix cfg

* fix logger
* fix trainer for int64 on windows

* polish trainer

* new usage of amp training.

* change the usage of amp and pure fp16 training.

* modified code as reviews

* add static running in dygraph

polish codes and docs to adapt Paddle2.0-rc