MarkAliasesPrepare to recognize meta ops with DID loop split.

[wujingyue]

See https://github.com/NVIDIA/Fuser/tree/bug3902 for a repro.

The root cause is having too many assumptions on allocation being a permutation of logical, e.g.,

Fuser/csrc/alias_analysis.cpp

Line 138 in bd2c81d

if (!ir_utils::computePermutation(

cc @jjsjann123: have you thought about this? Wonder whether/how ID model can play a role here.

[jjsjann123]

very roughly looking at the analysis,

In function AliasFinder::mapInLayoutToOutRoot

  std::unordered_map<IterDomain*, IterDomain*> in_logical_to_out_root =
      PairwiseLogicalDomainMap(in, out).mapProducerToConsumer();

Instead of using the PairwiseLogicalDomainMap, If we switch to id_model, we can try to map it with ExactGraph.

So inside Layout, instead of storing std::vector<IterDomain*>, I think we can do std::vector<ValGroup>?

So we still won't be able to support transforms, but for the repro case, where the meta operation just preserves the allocation domain from producer, we should be able to identify them as the same.

[jjsjann123]

So we still won't be able to support transforms

Even for transforms, if id_model can figure out the equivalence among IterDomain through transforms, we would benefit from that as well I guess?

For TV's without an allocation domain, we'll need to figure out if/how to replay the transformation. I don't know if we already have utility that does that. But seems like something worth exploring.