Implement Fusion definition equality check

[rdspring1]

This PR creates a check to determine if two separate Fusion objects have the same definition. This check is unlike existing equality IR checks the enforce that the object must be exactly the same. This second part of creating an LRU Cache for direct bindings.

sameDefinition details

Val::sameDefinition checks the Val's definition. Expr::sameDefinition checks the Expr's input arguments.
Fusion::sameDefinition starts from each output and traverses the entire Fusion DAG. Unlike the legacy TensorRecord check,
TensorView::sameDefinition directly compares the logical and allocation domain, which handles non-trival operations like
preprocessGroupedMatmulInputSf.

Differences between Val::sameAs and Val::sameDefinition

• Val::sameVal contains common code for Val::sameAs and Val::sameDefinition
• sameAs determines if a Statement has the exact same result as the other value.
• For sameAs, two values without definition nor value are not equal. This is used in concretization.
  In addition, if either Val has a non-deterministic definition, then the Vals are not the same.
• sameDefinition determines if a two values will create the same fusion definition.
• Symbolic values and NaNs are considered the same. Also, the value's integer name is checked to ensure argument order.

Differences between Expr::sameAs and Expr::sameDefinition

• Expr::sameDefinition skips the isDeterministic.

PR Stack

• Implement simple hash function for Fusion #5453
• Implement Fusion definition equality check #5454 <<< This PR.
• Create LRU cache for direct bindings. #4893

[github-actions]

Review updated until commit 9f6ee7c

Description

• Implement sameDefinition for Fusion to compare structural equivalence

• Add sameDefinition methods across IR nodes: Val, Expr, IterDomain, etc.

• Support alias and symbolic value handling in definition equality checks

• Introduce comprehensive tests for fusion definition equality

---

Changes walkthrough 📝

	Relevant files
Enhancement	8 files fusion.cpp Add Fusion::sameDefinition for structural comparison          +46/-0    base_nodes.cpp Implement Val::sameVal and sameDefinition methods                +90/-28  nodes.cpp Add sameDefinition for IterDomain and TensorDomain              +93/-39  tensor_view.cpp Implement TensorView::sameDefinition override                        +15/-0    fusion.h Declare Fusion::sameDefinition method                                        +3/-0      base_nodes.h Declare sameDefinition and sameVal in Val and Expr              +15/-0    interface_nodes.h Declare TensorView::sameDefinition override                            +2/-0      internal_base_nodes.h Declare sameDefinition for IterDomain and TensorDomain      +4/-0
Tests	1 files test_fusion_hash.cpp Add tests for Fusion sameDefinition and hashing                    +93/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review Input Order Sensitivity The sameDefinition function in Fusion checks inputs and outputs by index, implying order sensitivity. This may lead to false negatives when comparing functionally equivalent fusions with reordered inputs/outputs. Consider whether definition equality should be order-independent. if (inputs().size() != other.inputs().size()) { return false; } if (outputs().size() != other.outputs().size()) { return false; NaN Handling in sameDefinition The sameDefinition method handles NaN values specially for double types, but this logic may need to be extended to other floating-point types (e.g., float) for consistency. Consider whether this NaN equality logic should be applied more broadly. if (value_.is<double>() && std::isnan(value_.as<double>()) && std::isnan(other_val->value_.as<double>())) { return true; Complex Allocation Domain Check The strideOrder method returns empty for complex allocation domains, which affects how tensor domains are compared in sameDefinition. This special case should be validated to ensure it doesn't lead to incorrect equality determinations between tensor domains that should be considered different. bool is_complex_allocation = std::all_of( allocation_domain_.begin(), allocation_domain_.end(), [](IterDomain* id) { return id->extent()->definition() != nullptr; }); if (is_complex_allocation) { return {}; }