[Relay][TF] Keep node name in span

[lixiaoquan]

With this PR, it is easy to trace a relay CallNode to its source in an imported model. It looks like:

%57 = nn.relu(%56) /* resnet_model/Relu_9 */;
%58 = transpose(%resnet_model/conv2d_12/kernel, axes=[3, 2, 0, 1]);
%59 = nn.conv2d(%57, %58, padding=[0, 0, 0, 0], channels=128, kernel_size=[1, 1]) /* resnet_model/conv2d_12/Conv2D */;
%60 = nn.batch_norm(%59, %resnet_model/batch_normalization_10/gamma, %resnet_model/batch_normalization_10/beta, %resnet_model/batch_normalization_10/moving_mean, %resnet_model/batch_normalization_10/moving_variance, epsilon=1.001e-05f) /* resnet_model/batch_normalization_10/FusedBatchNorm */;

I'm not sure whether it is the correct to use span.SourceName to store node name, but line, column doesn't make sense for an imported model.

@jroesch @junrushao1994 @zhiics Could you please help to review? Thanks

[FrozenGene]

Ah...I think it is a nice feature and should enable it in all frontend, this is good for profiling (like we use debug graph runtime to see every layer's output). But I remember @tqchen said we will have another WIP feature to do this?

[tqchen]

I now think that span is the right way to do it, as long as we don't rely the information in passes.

[zhiics]

LGTM. I think this is also what @jroesch was adding.

[tqchen]

Perhaps we could also do similar things for other frontends

[jroesch]

@lixiaoquan I think we can do this, but I am not sure if overloading the source spans are the right way? perhaps we can sub-class or modify spans for handling whether they point into a source file or into an imported graph? I think it might be good to clarify this design.

[lixiaoquan]

@jroesch I agree with you. It looks like a makeshift to use SourceName to store node name, I did this because it introduced very little change. I think we can add a hint field. I added a thread https://discuss.tvm.apache.org/t/expand-span-for-imported-module/8435