Remove validated ORT core patches

[tzuhsuanwei]

Summary

This PR removes ORT-core patches 0002, 0004, and 0005 so the QNN plugin can build against a cleaner, unmodified upstream ORT source tree.

Patch 0002 previously changed ORT core solely to prevent the model runner from executing three QNN HTP RoiAlign test directories. The upstream plugin runner accepts a suite directory but has no per-case skip argument. Removing 0002 without replacement would make the normal full node-suite workflow execute est_roialign_mode_max, which is currently unsupported on QNN HTP and returns Not implemented.

Therefore this PR moves that test-selection policy out of ORT core and into a QNN-owned Python filter. Windows and Linux launchers call the filter before invoking the upstream plugin runner; Appium/QDC applies it before uploading the node suite to the device. The runner receives a temporary filtered suite, so ORT itself remains pristine and no longer needs a QNN-specific skip patch. This preserves the prior test workflow while keeping the ORT patch chain clean.

• Remove 0002, 0004, and 0005 from the FetchContent patch chain and delete their patch files.
• Re-home 0002's former RoiAlign exclusions in the QNN test harness, not ORT core.

Final patch disposition

Patch	Decision	Evidence / removal gate
0001 plugin EP runner	Keep	QNN still depends on the local plugin-EP runner implementation.
0002 RoiAlign HTP exclusions	Remove	Upstream runner has no individual-case skip CLI. QNN filter explicitly excludes the three former cases before runner invocation and logs reasons.
0003 arm64ReproDir	Keep	Upstream arm64x.cmake unconditionally overwrites arm64ReproDir; outer QNN CMake cannot fix it after ORT configures. Keep until upstream supplies a cache variable/override hook and ARM64X repro paths validate.
0004 monolithic LSTM option	Remove	Plugin runner forwards arbitrary key/value provider options; explicit enumeration is redundant.
0005 MSVC C4875	Remove	Pinned ORT v1.29 contains upstream GSL C4875 root-cause fix.
0006	No action	No patch file or application reference exists.
0007 Android QNN_LIB_FILES copy	Keep	Android Java/AAR path can still invoke cmake -E copy with an empty source list.
0008 rank-5/6 HandleReshapeSplit	Keep	QNN SpaceToDepth/ChannelShuffle fusions need hardening for the optimizer-produced form.

Validation

• Fresh ARM64 source build after moving prior build tree aside.
• Fresh ort_core patch stamp omits 0002; fresh ORT TestCase.cc/main.cc contain no RoiAlign exclusions.
• Fresh ARM64 runner and matching DLLs deployed to V81.
• V81 HTP run with backend_type|htp and htp_arch|81: Models 1, Succeeded 1, Not implemented 0, Failed 0.
• Filter staging suite excludes test_roialign_aligned_false, test_roialign_aligned_true, and test_roialign_mode_max while retaining test_abs.
• Python 3.8 filter smoke test passed; Windows launcher PowerShell syntax passed.
• Repository lint was run before the PR reduction/reinstatement; this PR's restored filter code is the same validated implementation.

[qti-chuteng]

The exclusion reasons describe these as QNN HTP limitations, but the node suite is currently executed with the QNN CPU backend in run_tests.sh, run_tests.ps1, and test_ort.py.

Because filter_model_test_suite() does not receive a backend, these cases are removed for every backend. This may silently reduce valid CPU coverage if the limitation is actually HTP-specific.

Could we clarify the intended scope?

• If this is a QNN EP-wide limitation, please change the reasons from “QNN HTP” to “QNN EP”.
• If it is HTP-only, please include the backend in the exclusion key/signature and add CPU-vs-HTP tests.

[tzuhsuanwei]

Thanks. I validated the three RoiAlign opset-22 cases on V81 and made the exclusions backend-aware. With backend_type|cpu, all three cases report Not implemented. With HTP, roialign_aligned_false and roialign_aligned_true succeed, while roialign_mode_max remains unsupported. The filter now receives the backend from the Windows, Linux, and QDC/Appium call sites: CPU excludes the three unsupported opset-22 cases, while HTP excludes only mode=max. Unit coverage now verifies CPU, HTP, and unconfigured-backend behavior.

[qti-chuteng]

Could we guarantee cleanup of filtered_test_root on every exit path?

With set -e, a failure in model_test_filter.py, an interrupted runner, or another early return can exit before line 78 and leave the copied suite under the system temporary directory.

Please register cleanup immediately after mktemp, for example with a function-scoped RETURN/EXIT trap, and use rm -rf -- "${filtered_test_root}". A failure-path test would also help
prevent this from regressing.

[tzuhsuanwei]

Done. Cleanup is registered immediately after mktemp with RETURN and EXIT traps and uses rm -rf --. The filter-failure path is handled explicitly, and cleanup failures cannot replace the model-runner result. I also added a Linux regression test that verifies cleanup on both a normal function return and an early shell exit.

[qti-chuteng]

Please wrap creation and use of $FilteredTestPath in try/finally.

At present, a terminating error from the filter or runner bypasses line 178, leaving a partially copied model suite in the system temp directory. This is especially likely for disk-full,
permission, or interrupted-job failures.

The finally block should remove the directory with -LiteralPath, while preserving the runner result for the existing failure handling.

[tzuhsuanwei]

Done. Creation and use of the filtered suite are wrapped in try/finally. The finally block removes the directory with Remove-Item -LiteralPath -Recurse -Force. The runner result is captured before cleanup, and a cleanup failure emits a warning rather than overriding the existing test result.