Should we test `-X lazy_imports=all` more?

Bug report

Bug description:

Inspired by #149321 I decided to see how well the test suite works with PYTHON_LAZY_IMPORTS=all / -X lazy_imports=all, which makes all imports lazy. It's nowhere near as bad as the none option that we're considering removing, but a few parts of the stdlib are broken, and it seems worth considering if we want to fix these.

Concrete issues I noticed:

• #144957 breaks test_typing
• test_unittest fails because unittest.main becomes a module instead of a function.
• test_struct is broken because of a test that does exec("import struct") in a function; lazy imports are disallowed within functions. Similar issues affect test_enum, test_traceback, and test_future.
• test_symtable is broken because it fills its _flags list by iterating over globals(), which means it starts containing lazy import marker objects. Similarly test_inspect breaks because it iterates over vars() of a module.

A few tests fail for legitimate reasons; test_compileall tests that an import triggers pyc compilation; test_builtin asserts that __import__ is used when in fact __lazy_import__ gets used.

Full list of failed modules on my machine:

test.test_future_stmt.test_future
test.test_inspect.test_inspect
test.test_pydoc.test_pydoc
test___all__
test__interpreters
test_builtin
test_capi
test_clinic
test_compileall
test_crossinterp
test_datetime
test_enum
test_generated_cases
test_idle
test_import
test_importlib
test_interpreters
test_json
test_lazy_import
test_struct
test_subprocess
test_symtable
test_tools
test_trace
test_traceback
test_tracemalloc
test_typing
test_unittest
test_xmlrpc
test_zipfile

CPython versions tested on:

CPython main branch

Operating systems tested on:

macOS

Linked PRs

• gh-149739
• gh-151090
• gh-151105

Comece revisando os PRs vinculados gh-149739, gh-151090 e gh-151105 e, em seguida, reproduza as falhas com PYTHON_LAZY_IMPORTS=all ou -X lazy_imports=all. Investigue os testes stdlib mencionados, separando as falhas causadas por lazy imports dos testes que verificam um comportamento intencional. Considera-se concluído quando o escopo estiver acordado e os testes relevantes forem corrigidos ou suas falhas esperadas forem documentadas.