[WIP][E2E]Fix e2e errors

[Xinyu302]

1. When we run affine-opt in the e2e pipeline, it will fail at fold-memref-alias-ops pass.

module {
  func.func @fold_static_stride_subview_with_affine_load_store(%arg0: memref<12xf32>, %arg1: index, %arg2: index, %arg3: index, %arg4: index) {
    %c0 = arith.constant 0 : index
    %c1 = arith.constant 1 : index
    %c10 = arith.constant 10 : index
    scf.for %arg5 = %c0 to %c10 step %c1 {
      %subview = memref.subview %arg0[%arg5] [1] [1] : memref<12xf32> to memref<f32, strided<[], offset: ?>>
      %0 = affine.load %subview[] : memref<f32, strided<[], offset: ?>>
      vector.print %0 : f32
    }
    return
  }
}

Because we generate affine.load op inside scf.for op, and it is not allowed due to scf.for has no AffineScope trait. So we do lower-affine earlier to avoid this situation.
2. Fix SetSpaceOptPipeline for reduction codegen by adding "inline" and "gpu-launch-func-to-byre" pass
3. Still some bugs in reduction codegen in Resnet18/Whole.