Bazel can overwrite __init__.py with empty file in source tree

### Description of the bug:

If I mess up my BUILD structure, and forget to include `srcs = ["__init__.py"]`, it seems it can happen that Bazel overwrites the existing `__init__.py` with an empty file.

### Reproduce:

Setup the directory:
```
cd $(mktemp -d)
cat > MODULE.bazel <<-EOF
module(name = "test_module")

bazel_dep(name = "rules_python", version = "1.4.1")
EOF
mkdir some_package
cat > some_package/BUILD.bazel <<-EOF
load("@rules_python//python:defs.bzl", "py_library", "py_test")
py_library(
name = "_some_library",
srcs = ["_some_library.py"],
)
py_library(
name = "some_package",
srcs = ["__init__.py"],
deps = [":_some_library"],
)
py_test(
name = "correct_test",
srcs = ["test.py"],
main = "test.py",
deps = [":some_package"],
)
py_test(
name = "bad_dep_test",
srcs = ["test.py"],
main = "test.py",
deps = [":_some_library"],
)
EOF
cat > some_package/__init__.py <<-EOF
import some_package._some_library
method=_some_library.method
EOF
cat > some_package/_some_library.py <<-EOF
def method():
pass
EOF
cat > some_package/test.py <<-EOF
import some_package
some_package.method()
EOF
```

Double check the with `tree`. This gives:
```
.
├── MODULE.bazel
└── some_package
├── BUILD.bazel
├── __init__.py
├── _some_library.py
└── test.py
```

Run the good test -- it will pass.
```
bazel test //some_package:correct_test
```

Then run the bad test. It fails and overwrites `__init__.py`.
```
bazel test //some_package:bad_dep_test
```

The run the good test. It now fails.
```
bazel test //some_package:correct_test
```

Note: Setting common --incompatible_default_to_explicit_init_py in .bazelrc fixes this.

See also https://github.com/bazel-contrib/rules_python/issues/2945.

Riproduci il comportamento descritto nell’issue usando MODULE.bazel e some_package/BUILD.bazel, quindi analizza come i due target py_test gestiscono some_package/__init__.py. Esegui correct_test, bad_dep_test e poi di nuovo correct_test per confermare la sovrascrittura; il lavoro è completato quando il test errato non svuota più il file sorgente e il test corretto continua a superare.

Scritto dal modello di indicizzazione a partire dal testo della issue.