PubSub subscriptions automatically created in Python DirectRunner not reliably cleaned up

In the Python DirectRunner, we attempt to garbage-collect automatically-created PubSub subscriptions here (https://github.com/apache/beam/blob/446f311746b74bdfb420ba71d76097d683fde9a1/sdks/python/apache_beam/runners/direct/transform_evaluator.py#L370). However, this mechanism may not succeed, since it is called on the __del__ method when other references may have already been garbage-collected. We should further investigate the mechanism here and attempt to make it more reliable.

Imported from Jira [BEAM-2988](https://issues.apache.org/jira/browse/BEAM-2988). Original Jira may contain additional context.
Reported by: ccy.

Start in sdks/python/apache_beam/runners/direct/transform_evaluator.py around line 370 and inspect how automatically-created PubSub subscriptions are garbage-collected from __del__. Review the imported Jira issue for context, then investigate a cleanup mechanism that remains reliable when other references have already been collected; done means these subscriptions are consistently cleaned up.

Written by the indexing model from the issue text.