Fix incremental issue with namespace packages (option 1) (#18907)

Fixes #12664

A root cause is there is this stateful `_update_ns_ancestors` thing in
`modulefinder`, so if things get called in the wrong order, you can get
incorrect results.

See also the logic in `all_imported_modules_in_file` where we've fixed
several bugs like this previously, like #13124 and #10937

As a result of (seemingly accidentally) reusing imports across modules,
we can end up in a situation where the namespace gets added as a
dependency to all other modules and so on the cached run we attempt to
find namespace before package, which does not work

I am not sure this `imports` code path is even needed, so I will open an
alternate PR, see #18908.

Relevant history:
- #6582
- #6179

I can't write a good test for this because it requires something in
site_packages, but here's a minimal repro:
```
set -eux
rm -rf repro
mkdir repro
cd repro

SITEPACK=env/site-packages
mkdir -p $SITEPACK

mkdir $SITEPACK/ruamel
mkdir $SITEPACK/ruamel/yaml

printf 'from ruamel.yaml.main import *' > $SITEPACK/ruamel/yaml/__init__.py
printf 'import ruamel.yaml' > $SITEPACK/ruamel/yaml/main.py
printf '' > $SITEPACK/ruamel/yaml/py.typed

printf 'import ruamel.yaml' > a.py
printf 'import a' > main.py

rm -rf .mypy_cache
PYTHONPATH=$SITEPACK mypy main.py
PYTHONPATH=$SITEPACK mypy main.py
```

Author: hauntsaninja

mypy/semanal_main.py

@@ -389,6 +389,7 @@ def semantic_analyze_target(
     analyzer.global_decls = [set()]
     analyzer.nonlocal_decls = [set()]
     analyzer.globals = tree.names
+    analyzer.imports = set()
     analyzer.progress = False
     with state.wrap_context(check_blockers=False):
         refresh_node = node