[LLVMGPU] Combine parallel and reduction padding in LLVMGPUPadAndVectorDistribute

[Groverkss]

Since https://github.com/iree-org/iree/pull/18748 tensor.pad can be fused in with tiling. This patch combines the parallel and reduction padding passes into a single pass that pads at once, and the pads are later fused during tiling.