Add batch-level processing pipeline with xmap_reader for acceleration.

[xinghai-sun]

Resolve 2nd part of #230

DS2 的 _padding_batch 函数以 pipeline方式优化（此前仅instance preprocess 以 pipeline并行化，但尚未在batch 层面做)

[lispc]

虽然我没有实测证据，但是经验上略担心用线程而不是进程可能速度overhead更大。
以前用的一个Tensorflow上层库使用 多进程 + zeromq传递数据，做到了直接磁盘读jpg（而不是recordio格式）来训练，还是挺惊艳的。

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}