feat: merge windowsml usage using official example

[xieofxie]

• remove class WinML as it is confusing
• also incidentally fixes a bug introduced by fix: resolve_device/resolve_eps uses get_registered_ep_devices #712 — the two independent singletons (WinML and WinMLEPRegistry) both call ort.register_execution_provider_library(...) but dedup separately, so after fix: resolve_device/resolve_eps uses get_registered_ep_devices #712 winml perf --device npu emits a library is already registered under QNNExecutionProvider traceback (benign but alarming, swallowed by except Exception)

[timenick]

Quick question on the switch to with EpCatalog() as catalog: — the old code at ep_registry.py:92-104 had a long workaround comment explaining that WinMLEpCatalogRelease (called from EpCatalog.close() / __del__) crashes with ACCESS_VIOLATION (0xC0000005) on some QNN NPU driver configurations — a Windows SEH exception that Python try/except cannot catch. The workaround was to null out self._catalog._handle = None so close() becomes a no-op.

The new with EpCatalog() as catalog: still invokes close() via __exit__, which puts that crash path back — just relocated from interpreter shutdown to the return of _load_ep_catalog.

Two questions:

1. Has the underlying WinMLEpCatalogRelease issue been fixed upstream in windowsml? The original TODO said "Remove once windowsml fixes WinMLEpCatalogRelease to be safe during process teardown."
2. If not, has winml perf --device npu been verified on a QNN NPU machine after this change to confirm the crash doesn't return?

Side note: this PR also incidentally fixes a bug introduced by #712 — the two independent singletons (WinML and WinMLEPRegistry) both call ort.register_execution_provider_library(...) but dedup separately, so after #712 winml perf --device npu emits a library is already registered under QNNExecutionProvider traceback (benign but alarming, swallowed by except Exception). Worth calling out in the PR title/description so it shows up in git log --grep later.

Generated with Claude Code

[xieofxie]

Quick question on the switch to with EpCatalog() as catalog: — the old code at ep_registry.py:92-104 had a long workaround comment explaining that WinMLEpCatalogRelease (called from EpCatalog.close() / __del__) crashes with ACCESS_VIOLATION (0xC0000005) on some QNN NPU driver configurations — a Windows SEH exception that Python try/except cannot catch. The workaround was to null out self._catalog._handle = None so close() becomes a no-op.

The new with EpCatalog() as catalog: still invokes close() via __exit__, which puts that crash path back — just relocated from interpreter shutdown to the return of _load_ep_catalog.

Two questions:

1. Has the underlying WinMLEpCatalogRelease issue been fixed upstream in windowsml? The original TODO said "Remove once windowsml fixes WinMLEpCatalogRelease to be safe during process teardown."
2. If not, has winml perf --device npu been verified on a QNN NPU machine after this change to confirm the crash doesn't return?

Side note: this PR also incidentally fixes a bug introduced by #712 — the two independent singletons (WinML and WinMLEPRegistry) both call ort.register_execution_provider_library(...) but dedup separately, so after #712 winml perf --device npu emits a library is already registered under QNNExecutionProvider traceback (benign but alarming, swallowed by except Exception). Worth calling out in the PR title/description so it shows up in git log --grep later.

Generated with Claude Code

from my understanding, this is the official usage of windowsml lib https://pypi.org/project/windowsml/ I feel it should be well tested so that access violation should not exist