Request ConvertToSignlessPass in Stablehlo

[qingyunqu]

Request description

When I use stablehlo-to-linalg in torch-mlir(https://github.com/llvm/torch-mlir/pull/3367), I find that it will emit builtin.unrealized_conversion_cast in the following case:

module attributes {torch.debug_module_name = "Uint8ToFloat"} {
  func.func @forward(%arg0: tensor<3xui8>) -> tensor<3xf32> {
    %0 = stablehlo.convert %arg0 : (tensor<3xui8>) -> tensor<3xf32>
    return %0 : tensor<3xf32>
  }
}

#map = affine_map<(d0) -> (d0)>
module attributes {torch.debug_module_name = "Uint8ToFloat"} {
  func.func @forward(%arg0: tensor<3xui8>) -> tensor<3xf32> {
    %0 = builtin.unrealized_conversion_cast %arg0 : tensor<3xui8> to tensor<3xi8>
    %1 = tensor.empty() : tensor<3xf32>
    %2 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%0 : tensor<3xi8>) outs(%1 : tensor<3xf32>) {
    ^bb0(%in: i8, %out: f32):
      %3 = arith.uitofp %in : i8 to f32
      linalg.yield %3 : f32
    } -> tensor<3xf32>
    return %2 : tensor<3xf32>
  }
}

I think I need mhlo's ConvertToSignlessPass to eliminate the builtin.unrealized_conversion_cast.

[GleasonK]

Hello! If you have time to migrate that pass to StableHLO, we're happy to house it. Looking at the implementation, there is nothing specific to MHLO about that pass so it likely can be shared by both StableHLO/MHLO.