[fusion] Fold unit dims of globals

[MaheshRavishankar]

During compilation of quantized SDXL model there are artifacts of this form

  %extracted_slice_219 = tensor.extract_slice %236[0, 0, 0] [2, 4096, 2560] [1, 1, 1] : tensor<2x4096x5120xf16> to tensor<2x4096x2560xf16>
  %extracted_slice_220 = tensor.extract_slice %236[0, 0, 2560] [2, 4096, 2560] [1, 1, 1] : tensor<2x4096x5120xf16> to tensor<2x4096x2560xf16>
  %expanded_221 = tensor.expand_shape %extracted_slice_219 [[0], [1], [2, 3, 4]] output_shape [2, 4096, 1, 1, 2560] : tensor<2x4096x2560xf16> into tensor<2x4096x1x1x2560xf16>
  %expanded_222 = tensor.expand_shape %extracted_slice_220 [[0], [1], [2, 3, 4]] output_shape [2, 4096, 1, 1, 2560] : tensor<2x4096x2560xf16> into tensor<2x4096x1x1x2560xf16>
  %237 = tensor.empty() : tensor<2x4096x1x1x2560xi8>
  %238 = flow.dispatch.region -> (tensor<2x4096x1x1x2560xi8>) {
    %5295 = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>, affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>, affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d4)\>, affine_map<(d0, d1, d2, d3, d4) -> ()>, affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} ins(%expanded_221, %expande\d_222, %__auto.down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.premul_input, %__auto.down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.q_input3Ascale : tensor<2x4096x1x1x2560xf16>, tensor<2x40\96x1x1x2560xf16>, tensor<1x1x2560xf16>, tensor<f32>) outs(%237 : tensor<2x4096x1x1x2560xi8>) {

Here we would like to fuse the extract_slice with its consumer. That gets blocked due to the presence of the tensor.expand_shape in between the slice and its use in the dispatch. The real reason why these epxand shapes exist is because after FoldUnitDims, the unit dimensions in the globals like %__auto.down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.premul_input and %__auto.down_blocks.1.attentions.0.transformer_blocks.0.ff.net.2.q_input3Ascale dont get folded away. This results in the collapse_shapes not fully folding away, which later propagation passes pick up.
While this could be accounted for during the propagation passes, it is also worth just folding the unit dimensions in the global variables away.