WAR for supporting the rotation + residual pattern

[naoyam]

Stacked on top of #3549. This is also a WAR for #3455 and necessary to schedule RoPE-like rotation patterns.

Because of the issue, a tensor may have two IDs that are exactly mapped. For example, when an ID is sliced to half and then is padded back to the same size, and the final output ID is used with the initial input ID, the initial input and the final output IDs get mapped together. This can make it difficult to use scheduleLoopDomainsLike. For example, if a reference has a split that is done with the final output ID, and we want to replay the split on other tensors, it becomes ambiguous whether the split is done with the initial input or the final output since both are exactly mapped.

To avoid this ambiguity, this PR adds a flag to indicate that we just want to update the current loop domain with a reference domain. As seen in the added tests, this flag is used to propagate the scheduling of a reference tensor once all resize ops are propagated to inputs. Specifically, the overall scheduling follows this pattern:

1. Propagate all slice and pad ops to fusion inputs
2. Pick and schedule a reference tensor
3. Propagate the scheduling of the reference tensor to the other tensors

scheduleLoopDomainsLike with the flag is used at step 3. For that step, we know that we don't need to schedule each tensor with a complex replay path, like some backward ops followed by some other forward ops, but we just need to update the current loop domain by replaying the diff with the reference domain.

[naoyam]

!test

[naoyam]

!build

[jacobhinkle]

LGTM. Only comment is minor about naming of the param update_mode.

[naoyam]

!build