Enable QNN unit tests on Windows ARM64 internal-symbol builds

[tzuhsuanwei]

Summary

Enable the QNN internal-symbol unit-test build on Windows ARM64.

This stacks on #831 and adds the ARM64 build path plus a Draft-PR workflow: build on an x64 Windows builder, then run the QNN unit suite on a Windows ARM64 test runner.

CI scope

• Windows ARM64 only
• runs the existing mainline QnnUnit_* component/unit suite
• supports an optional supplied snapshot-golden artifact for manual validation
• no Linux jobs or shared golden-store publishing changes

Windows ARM64 test coverage

The original ARM64 CI result was 861 tests: 858 passed, 3 skipped. After the changes below, the pre-rebase Windows ARM64 CI result was 863 tests: 861 passed, 2 skipped.

• SetupBackend_HTP_WithHtpArch_ReportsPlatformArch

  • Keeps the configured V73 assertion on x86, where the HTP emulator reports the configured arch.
  • On ARM64, verifies that GetHtpArch() returns a discovered physical HTP arch (!= NONE) instead of assuming V73.

• ValidateCompatibilityInfo_MalformedOnArm64_LogsDeserializationFailureAndReturnsNotApplicable

  • Adds an ARM64 parser-path test for malformed compatibility info.
  • Verifies that ARM64 reaches deserialization and returns EP_NOT_APPLICABLE, rather than taking the x86 platform-skip path.

• HTP host device discovery

  • Retains the existing CPU-host case as an ARM64 skip because the CPU QNN backend is only enabled on the x86 coverage host.
  • Adds GetSupportedDevices_HtpHost_CreatesEpDevice for ARM64, verifying that a Qualcomm NPU/HTP host device creates an EP device.

Rebase, Accuracy, and Snapshot validation

This branch includes the #831 Phase 3 tier infrastructure (160e747423), which enables QNN_EP_ACCURACY_UT for internal-symbol builds.

The supplied-golden validation (run 36551637367) passed:

• The Accuracy listing contains all 18 current Clip Accuracy cases in four suites: QnnUnit_Clip_AccuracyTest (7), QnnUnit_Clip_Accuracy_QDQFloatTest (5), QnnUnit_Clip_Accuracy_QDQQuantTest (4), and QnnUnit_Clip_Accuracy_FoldedConstTest (2).
• All 18 Accuracy cases ran in the Windows ARM64 tier gate; none were skipped.
• The supplied golden artifact was downloaded successfully. All 18 Snapshot / SessionSnapshot cases were executed (2 SessionSnapshot and 16 Snapshot), and all reported QNN_SNAPSHOT_DRIFT; none were QNN_GOLDEN_ABSENT.
• Snapshot drift is retained in the CI log and the remaining tier run validates correctness through Accuracy / Component / Integration: 900 passed, 0 failed.

The supplied golden was produced from a different source baseline, so this result validates the drift-to-Accuracy route rather than snapshot equality. A Windows ARM64 golden generated from the matching source/config is required before the 18 snapshot comparisons can be expected to match exactly.