mlir CollapseShapeOp ExpandShapeOp的构造

作用
CollapseShapeOp把一些维度合并，总Tensor的size不变；ExpandShapeOp把一些维度拆开，总Tensor的size不变。
create方式

SmallVector<int64_t> rhsShapeSimplified；
auto targetRhsType = RankedTensorType::get(rhsShapeSimplified, cast<RankedTensorType>(rhsType).getElementType());
rhsToRewriter =rewriter.create<tensor::CollapseShapeOp>(op->getLoc(), targetRhsType(目标type), op->getOperand(1)(原始value), rhsReassociation);

Reassociation: 输入Tensor到输出Tensor对应的维度坐标信息：把哪些维度合并在一起
rhsReassociation类型是SmallVector本质是SmallVector<SmallVector<int64_t>>

tensor<1x1x1x4x3xf16> reassociation: [[0, 1, 2, 3], [4]]; 
tensor<1x3x4xf16> reassociation: [[0, 1], [2]]auto targetOutType = RankedTensorType::get(outShape, cast<RankedTensorType>(outType).getElementType());lastToReplace = rewriter.create<tensor::ExpandShapeOp>(op->getLoc(), targetOutType, matmul, outReassociation);

Reassociation: 输出Tensor到输入Tensor对应的维度坐标信息：把哪些维度合并在一起

获取Reassociation的方式，可以直接调三方接口

std::optional<SmallVector<ReassociationIndices>>
mlir::getReassociationIndicesForCollapse(ArrayRef<int64_t> sourceShape,ArrayRef<int64_t> targetShape)