Fixed the memory bugs when the input vector for innerproduct has 3 dims (3D vector innerproduct)

[parrotsky]

Fixed the memory bugs when the input vector for innerproduct has 3 dims (3D vector innerproduct) on arm and x86. We collapse the 3d vector back to 2d vector.

[tencent-adm]

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.}