[WIP][mlir] DenseElementsAttr::reshape(arg): make arg a shape, not a type

[newling]

The motivation is that it would have prevented the issue detected in #147691.

[MengmSun]

I think if result type conversion is intended then we should keep the element type of converted result type and use for subsequent lowering. Clone with newType.getElementType instead of curType.getElementType.

[MengmSun]

Otherwise the converted information like fp8->i8 will lost.
For example,

func.func @canonicalize_extract_shapecast_different_element_type() -> vector<1x192xf8E4M3FN> {
      %0 = arith.constant dense<1.000000e+00> : vector<192xf8E4M3FN>
      %1 = vector.shape_cast %0 : vector<192xf8E4M3FN> to vector<1x192xf8E4M3FN>
      return %1 : vector<1x192xf8E4M3FN>
}

will be converted to as below after -convert-to-llvm -canonicalize="test-convergence"

module {
  llvm.func @canonicalize_extract_shapecast_different_element_type() -> !llvm.array<1 x vector<192xf8E4M3FN>> {
    %0 = llvm.mlir.constant(dense<1.000000e+00> : vector<1x192xf8E4M3FN>) : !llvm.array<1 x vector<192xf8E4M3FN>>
    llvm.return %0 : !llvm.array<1 x vector<192xf8E4M3FN>>
  }
}

However, it should be

module {
  llvm.func @canonicalize_extract_shapecast_different_element_type() -> !llvm.array<1 x vector<192xi8>> {
    %0 = llvm.mlir.constant(dense<1.000000e+00> : vector<1x192xf8E4M3FN>) : !llvm.array<1 x vector<192xi8>>
    llvm.return %0 : !llvm.array<1 x vector<192xi8>>
  }
}

[newling]

Hi @MengmSun -- I just tried your example above on this branch with mlir-opt -convert-to-llvm --canonicalize test.mlir and it gives

llvm.func @canonicalize_extract_shapecast_different_element_type() -> !llvm.array<1 x vector<192xi8>> {
  %0 = llvm.mlir.constant(dense<1.000000e+00> : vector<1x192xf8E4M3FN>) : !llvm.array<1 x vector<192xi8>>
  llvm.return %0 : !llvm.array<1 x vector<192xi8>>
}

which I think is what we want