Prevent cross-EP Cast fusion in RemoveDuplicateCastTransformer

[fs-eire]

Description

Fixes an incorrect Cast deduplication across execution provider boundaries.
Today, Cast(int64->float, CPU) -> Cast(float->float16, WebGPU) can be fused into Cast(int64->float16, WebGPU), which is invalid for WebGPU and can fail kernel lookup.

This change adds an EP check so Cast fusion only happens when both nodes are on the same EP, and adds a regression test for this scenario.

Motivation and Context

Fixes #27291

[Copilot]

Pull request overview

This PR fixes an incorrect Cast node fusion across execution provider (EP) boundaries in the RemoveDuplicateCastTransformer. The bug allowed consecutive Cast nodes assigned to different EPs to be fused, resulting in a node with an input type unsupported by its EP (e.g., Cast(int64->float, CPU) -> Cast(float->float16, WebGPU) being fused into Cast(int64->float16, WebGPU), which fails because WebGPU doesn't support int64 inputs).

Changes:

• Added EP boundary check in RemoveDuplicateCastTransformer to prevent cross-EP Cast fusion
• Added regression test verifying Cast nodes on different EPs are not fused

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/core/optimizer/insert_cast_transformer.cc	Added cross-EP check before fusing Cast nodes, with detailed comments explaining the issue and solution
onnxruntime/test/framework/insert_cast_transformer_test.cc	Added regression test that creates CPU->WebGPU Cast chain and verifies they remain separate after transformation

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.