Generalize the tensor.gather to linalg.generic

[zhczhong]

The tensor.gather is not bufferized and here add a pass to transform it to linalg.generic so that it could be further lowered correctly.(Also add tensor::populateDecomposeTensorConcatPatterns to handle tensor.concat)

func.func @gather_single_gather_dim(%arg0: tensor<2x2x2x2xf32>, %arg1: tensor<2x3x1xindex>) -> tensor<2x3x2x2x2xf32> {
tensor<2x3x1xindex>) outs(%0 : tensor<2x3x2x2x2xf32>)
  %1 = tensor.gather %arg0[%arg1] gather_dims([1]) : (tensor<2x2x2x2xf32>, tensor<2x3x1xindex>) -> tensor<2x3x2x2x2xf32>
  return %1 : tensor<2x3x2x2x2xf32>
}

to

  func.func @gather_single_gather_dim(%arg0: tensor<2x2x2x2xf32>, %arg1: tensor<2x3x1xindex>) -> tensor<2x3x2x2x2xf32> {
    %0 = tensor.empty() : tensor<2x3x2x2x2xf32>
    %1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%arg1 : tensor<2x3x1xindex>) outs(%0 : tensor<2x3x2x2x2xf32>) {
    ^bb0(%in: index, %out: f32):
      %2 = linalg.index 2 : index
      %3 = linalg.index 3 : index
      %4 = linalg.index 4 : index
      %extracted = tensor.extract %arg0[%2, %in, %3, %4] : tensor<2x2x2x2xf32>
      linalg.yield %extracted : f32
    } -> tensor<2x3x2x2x2xf32>
    return %1 : tensor<2x3x2x2x2xf32>
  }

[dchigarev]

Looks good overall, left a few questions regarding the logic

[kurapov-peter]

Does it work with dynamic shapes?

[zhczhong]

The dynamic support and the corresponding test is added