The new framework needs to allocate aligned memory.

[qingqing01]

When porting the LSTM operator, the AVX version of LSTM kernel failed. With the help of @hedaoyuan, we found it caused by the misaligned memory address.