Make patch_submodule robust to __builtins__ being a module (PyPy / __main__)

[vineethsaivs]

Closes #7636.

Problem

patch_submodule reads the builtins through globals()["__builtins__"]:

elif target_attr in globals()["__builtins__"]:  # if it's a builtin like "open"
    self.original[target_attr] = globals()["__builtins__"][target_attr]

The value of a module's __builtins__ global is a CPython implementation detail: it is the builtins module (not its dict) in the __main__ module and under PyPy, and only usually a dict in imported modules (docs). When it is the module, target_attr in globals()["__builtins__"] raises:

TypeError: argument of type 'module' is not iterable

so patching a builtin such as open fails. This is what #7636 hit running datasets streaming under PyPy.

Fix

Use builtins.__dict__ directly, which is always the builtins mapping regardless of the runtime. In the common CPython case (an imported module whose __builtins__ already is builtins.__dict__) the behavior is unchanged; it just also works under __main__/PyPy.

Test

test_patch_submodule_builtin_when_builtins_is_module sets datasets.utils.patching.__builtins__ to the builtins module (simulating the PyPy / __main__ runtime) and patches open. It raises the TypeError above on the current code and passes with this change; the existing patching tests are unaffected (full tests/test_patching.py is green).