Enable QNN unit tests on Linux ARM64 internal-symbol builds

[tzuhsuanwei]

Summary

Enable the QNN internal-symbol unit-test build on Linux ARM64 (HTP V68) without coverage instrumentation.

This adds a Linux ARM64 Draft-PR workflow that builds and runs the tier-first QNN flow, followed by the existing legacy QNN catch-all path.

CI scope

• Linux ARM64 only
• QnnUnit_* / QnnInteg_*, plus the legacy *Qnn* catch-all path
• no Windows jobs or shared golden-store publishing changes

Rebase, Accuracy, and Snapshot validation

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

The supplied-golden validation (run 36551633453) reused the Linux ARM64 build artifact and 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).
• On the real HTP V68 runner, Clip_int32 and all 9 QDQ Accuracy cases run and pass. The remaining 8 float/FP16-path cases are explicit V68 capability skips.
• 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 reported in the CI log and routes to the Accuracy gate. The non-snapshot tier run completed 914 passed, 11 skipped, 0 failed; legacy completed 1411 passed, 1269 skipped, 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 Linux ARM64 golden generated from the matching source/config is required before the 18 snapshot comparisons can be expected to match exactly.

Linux ARM64 coverage

• The internal-symbol build links onnxruntime_provider_test with the QNN EP and enables internal symbol access only in this test mode.
• All 19 Clip component tests are listed.
• All 9 QnnInteg_OrtApiTest cases run and pass. Conv and ConvTranspose use their existing float path on non-ARM64 and a V68-compatible U8 QDQ path on Linux ARM64.