jit benchmark use tensor

[tensor-tang]

Since tensor has alignment.

test=develop

[luotao1]

Since tensor has alignment.

Do you mean the memory alignment? Is the part solution of #15447?
Please add more descriptions in the first comment or create another issue to describe the problem.
And add comments in the codes, like To xxx, we use Tensor instead of std::vector<T>, please see issue:.

[tensor-tang]

This is not the solution of the issue, but just we need make the benchmark more credibility since we leant from the #15447 that some MKL env need memory allocated with alignment.