[Core] Add correctness tests for SpaceToDepth and MobileClip Attention fusion

[hariharans29]

Description

Add correctness tests for fusions introduced in #27883 and #27747. The tests introduced in those PRs only check if fusion went through but not if the fused nodes produced semantically right results as the unfused subgraphs. Adding those tests to prevent accidental breakage in case something changed in the fused node's backing kernel implementation.

Motivation and Context

Adress test coverage gap

[Copilot]

Pull request overview

This PR upgrades existing optimizer fusion tests to validate semantic correctness (fused vs. unfused outputs), not just that the fusion pattern triggered. It targets the SpaceToDepth fusion and the MobileClip Attention fusion added in prior PRs.

Changes:

• Convert SpaceToDepth fusion tests from TestGraphTransformer graph-structure-only checks to TransformerTester output comparisons plus transformed-graph assertions.
• Convert MobileClip Attention fusion tests to TransformerTester for correctness verification on CPU and CUDA (when available).
• Add small adapter helpers to run existing Status-based graph validators from TransformerTester’s void(InferenceSessionWrapper&) callback.

---

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

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.

---

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

[tianleiwu]

Reviewed the updated fusion tests against the current head.

The switch to TransformerTester improves these cases by checking unfused-vs-fused output equivalence while still asserting the expected transformed graph shape. I also checked the CUDA path carefully, and the new session-based flow is consistent with the current MobileCLIP fusion logic for unassigned nodes before partitioning.

I did not find any actionable issues in the current patch.