[Fix] Enhance robustness of DefuseOps

[ganler]

Another solution of fix regarding #8559

@merrymercy

[ganler]

Instead of playing tricks regarding variable names, we create a mapping (in the FunctionNode phase) from variable names to the corresponding expressions.

[YuchenJin]

Thanks @ganler! Do you know other passes that need to be modified to be compatible with AnnotateSpans due to the same root cause?

[ganler]

Thanks @ganler! Do you know other passes that need to be modified to be compatible with AnnotateSpans due to the same root cause?

@YuchenJin It is possible but I am not quite sure for now. I tried all pair combinations of the following passes, only found DefuseOps has such an issue.

    relay.transform.RemoveUnusedFunctions,
    relay.transform.Inline,
    relay.transform.PartitionGraph,
    relay.transform.ToGraphNormalForm,
    relay.transform.SimplifyInference,
    relay.transform.FoldConstant,
    relay.transform.AnnotateSpans,
    relay.transform.DefuseOps,
    relay.transform.FuseOps,
    relay.transform.SimplifyExpr,
    relay.transform.ToBasicBlockNormalForm,
    relay.transform.BatchingOps,
    relay.transform.AlterOpLayout,
    relay.transform.FoldScaleAxis,
    relay.transform.CanonicalizeOps,
    relay.transform.CanonicalizeCast,
    relay.transform.DeadCodeElimination,
    relay.transform.EliminateCommonSubexpr,
    relay.transform.CombineParallelConv2D,
    relay.transform.CombineParallelDense,
    relay.transform.CombineParallelBatchMatmul,
    relay.transform.FastMath,
    relay.transform.DynamicToStatic,
    relay.transform.FoldExplicitPadding,