* add parallel record decoder

* null data loader

* use real loader

* use libjpeg-turbo

* add streams support, TODO:1,disable internal buffer_;2,use template

* make persistent_in_stream support multiple files

* make compiler support new loader

* minor refine

* make new loader work on mnist

* update proto in benchmark and example

* refactor stream buffer filler

* refine persistent_in_stream

* workable

* add record_loader_op

* finish record loader op

* AddRecordLoaderOps

* make compiler work

* infer shape works

* add record_load_kernel

* let decode actor pass in_regst in normal way

* add kOFRecordPtr type

* remove record regst type

* change ALL_DATA_TYPE_SEQ to ALL_POD_DATA_TYPE_SEQ

* support OFRecordPtr blob

* complete decode_ofrecord_kernel

* allocate OFRecord in Blob<OFRecordPtr>

* fix of record ptr blob

* let actor manage the OFRecord blob

* let regst mgr own ofrecord memory

* workable

* remove useless code

* refine

* NormalRegst -> DataRegst

* OFRecord data type (#984)

* OFRecord data type

* placement new (#987)

* placement new

* fix

* remove useless code

* placement new OFRecord

* remove useless code

* Refactor stream (#985)

* refactor stream scanner

* let persistence_in_stream create the binary stream

* refine persistence_in_stream

* refine

* POD_DATA_TYPE_SEQ

* update placement proto in benchmark and example