[amoeba]

When CAdbcDriverImporter.Load is called with canUnload: true and the library doesn't export the requested entrypoint, the netstandard2.0 build (used on .NET Framework) frees the library twice: once before throwing and again in the finally block. That build uses its own NativeLibrary shim, whose GetExport returns IntPtr.Zero for a missing export instead of throwing, so it reaches a branch the net8.0/net10.0 builds never do. Freeing twice can unload a library that something else in the process still uses.

This PR adds a test for this case and leaves freeing the library to the finally block.

Bug found and PR created with Opus 5.5.

[amoeba]

CI should fail on Windows on a9edd1e.

[amoeba]

From https://github.com/apache/arrow-adbc/actions/runs/40781480559/job/123143975815?pr=4976

[xUnit.net 00:00:02.59]     Apache.Arrow.Adbc.Tests.CAdbcDriverImporterUnloadTests.Load_CanUnload_MissingEntrypoint_KeepsOtherReferences [FAIL]
  Failed Apache.Arrow.Adbc.Tests.CAdbcDriverImporterUnloadTests.Load_CanUnload_MissingEntrypoint_KeepsOtherReferences [227 ms]
  Error Message:
   The library was unloaded while another reference to it was still held
  Stack Trace:
     at Apache.Arrow.Adbc.Tests.CAdbcDriverImporterUnloadTests.Load_CanUnload_MissingEntrypoint_KeepsOtherReferences() in D:\a\arrow-adbc\arrow-adbc\csharp\test\Apache.Arrow.Adbc.Tests\CAdbcDriverImporterUnloadTests.cs:line 60

The error isn't a direct check but hopefully the way the test is structured makes this make sense.

[CurtHagenlocher]

Thanks!