[mlir][linalg] Convert tensor.pad to destination style

This can be a pre-processing for bufferization and allows for more efficient lowerings without an alloc.

Differential Revision: https://reviews.llvm.org/D142207