Add CPU-only MemcpyTransformer test with fake device EP

[GopalakrishnanN]

Description

Adds a CPU-only MemcpyTransformer test that exercises copy-node insertion at mixed-EP partition boundaries without requiring CUDA.

The new test introduces a tiny fake non-CPU execution provider whose default device is GPU-like, but it never executes kernels. The test:

• builds a CPU -> fake-device -> CPU graph;
• registers fake MatMul kernel metadata for the fake provider so MemcpyTransformer can inspect kernel metadata during graph rewrite;
• applies MemcpyTransformer;
• asserts a MemcpyFromHost is inserted before the fake-device node and a MemcpyToHost is inserted before the downstream CPU node.

This complements the existing CUDA-gated MemcpyTransformer tests with a locally runnable boundary-copy test.

Motivation and Context

EP partitioning/copy behavior is a high-impact performance/design area. Previously, the direct MemcpyTransformer tests were guarded by USE_CUDA, so CPU-only builds did not exercise this graph-rewrite logic. This test makes the core copy-insertion behavior verifiable without GPU hardware while still avoiding actual execution of fake-device kernels.

Testing

Built onnxruntime_test_all (Release, Windows) and ran:

.\onnxruntime_test_all.exe --gtest_filter="TransformerTest.MemcpyTransformerTestWithFakeDeviceProvider"

Result:

[==========] 1 test from 1 test suite ran.
[  PASSED  ] 1 test.

[Copilot]

Pull request overview

This PR adds a CPU-only unit test for MemcpyTransformer that validates copy-node insertion at mixed execution-provider boundaries without requiring CUDA. It does so by introducing a minimal fake non-CPU execution provider with a GPU-like OrtDevice, registering fake kernel metadata so MemcpyTransformer can query kernel placement info, and asserting that MemcpyFromHost/MemcpyToHost nodes are inserted at the CPU↔fake-EP boundaries.

Changes:

• Add a minimal FakeDeviceExecutionProvider (GPU-like device) for testing partition-boundary copy insertion in CPU-only builds.
• Add TransformerTest.MemcpyTransformerTestWithFakeDeviceProvider to validate MemcpyFromHost and MemcpyToHost insertion around a CPU → fake-EP → CPU graph.
• Refactor GetNotNullProviderPtrs helper so it’s available to the new non-CUDA-gated test.

[GopalakrishnanN]

DirectML follow-up: after building a DML-enabled shared test tree on this Intel Arc machine, I extended this PR with a real DML variant in addition to the fake-device test. Validation: built onnxruntime_test_all in �uild\Windows_DML18_Shared and ran TransformerTest.MemcpyTransformerTestWithFakeDeviceProvider:TransformerTest.MemcpyTransformerTestWithDmlProvider — 2/2 passed. Also ran DML smoke QuantizeLinearOpTest.Uint8_5D_NoZeroPoint_Opset21_DML — passed.

[github-actions]

You can commit the suggested changes from lintrunner.