[Intel] Add padding support to tensor descriptor

[whitneywhtsang]

It is implemented upstream in dba08d4.
This issue should add the support for Intel codegen, and revert 31c76e4.