How to reuse some existing op fusion capabilities in a new EP?

[alter-xp]

on InferenceSession::TransformGraph, the transformer order:

1. run level 1 optimizations. these only use ONNX operators.
2. partition nodes based on EP capabilities. EPs may fuse nodes during this process.
3. run all optimizations. level 2 and 3 optimizations use contrib ops.
4. insert cast nodes
5. insert copy nodes

In this order, since the level 2 optimization traversal is after the partitioning, if the EP needs to perform node fusion, it needs to implement it on its own. Is there a way to reuse existing op fusion in the EP?