fix(ThreadWorker): don't unload modules without __loader__ (backport)

DuckDB (and likely other packages) manually add their submodules (e.g. duckdb.duckdb.functional) during side-effects when the "parent" (e.g. duckdb.duckdb) is loaded. Since extension modules can't be reloaded, it is impossible to re-run these side-effects to recreate these modules. This commit preserves modules without the __loader__ attribute (which is added by Python's normal import machinery), sacrificing isolation for compatibility.

questionpy_server/worker/impl/thread.py

@@ -51,6 +51,13 @@ def run(self) -> None:
 
             sys.path = original_path
             for module_name in sys.modules.keys() - original_module_names:
+                if getattr(sys.modules[module_name], "__loader__", None) is None:
+                    log.debug(
+                        "Not unloading '%s', as it doesn't have a '__loader__' attribute and was probably added "
+                        "manually.",
+                        sys.modules[module_name].__name__,
+                    )
+                    continue
                 # Having reset the path, this forces questionpy and any package modules to be reloaded upon next import.
                 del sys.modules[module_name]