[TorchToLinalg] Incorrect lowering for simple dynamic collapse in aten.view

[zjgarvey]

The following IR is incorrectly lowered to linalg:

func.func @main_graph(%arg0: !torch.vtensor<[?,?,768],f32>) -> !torch.vtensor<[?,768],f32> {
  %int-1 = torch.constant.int -1
  %int768 = torch.constant.int 768
  %42 = torch.prim.ListConstruct %int-1, %int768 : (!torch.int, !torch.int) -> !torch.list<int>
  %43 = torch.aten.view %arg0, %42 : !torch.vtensor<[?,?,768],f32>, !torch.list<int> -> !torch.vtensor<[?,768],f32>
  return %43 : !torch.vtensor<[?,768],f32>
}

This converts to:

module {
  func.func @main_graph(%arg0: !torch.vtensor<[?,?,768],f32>) -> !torch.vtensor<[?,768],f32> {
    %0 = torch_c.to_builtin_tensor %arg0 : !torch.vtensor<[?,?,768],f32> -> tensor<?x?x768xf32>
    %int-1 = torch.constant.int -1
    %int768 = torch.constant.int 768
    %1 = torch.prim.ListConstruct %int-1, %int768 : (!torch.int, !torch.int) -> !torch.list<int>
    %c0 = arith.constant 0 : index
    %dim = tensor.dim %0, %c0 : tensor<?x?x768xf32>
    %c1 = arith.constant 1 : index
    %dim_0 = tensor.dim %0, %c1 : tensor<?x?x768xf32>
    %c2 = arith.constant 2 : index
    %c768 = arith.constant 768 : index
    %c-1_i64 = arith.constant -1 : i64
    %c768_i64 = arith.constant 768 : i64
    %cast = tensor.cast %0 : tensor<?x?x768xf32> to tensor<?x1x768xf32>
    %collapsed = tensor.collapse_shape %cast [[0], [1, 2]] : tensor<?x1x768xf32> into tensor<?x768xf32>
    %2 = torch_c.from_builtin_tensor %collapsed : tensor<?x768xf32> -> !torch.vtensor<[?,768],f32>
    return %2 : !torch.vtensor<[?,768],f32>
  }
}

It seems the lowering is unable to figure out that the back dim is untouched, and that the two dynamic dims are supposed to be flattened.

[zjgarvey]

I found this issue when debugging the IREE runtime failures for the model model--finetuned_distilgpt2_sst2_negation0.0001_pretrainedTrue_epochs1--jhaochenz in our onnx test suite.