This refactors the watch(), onChange() and AttachmentTest implementations to improve the reliability of the CI.

First, both watch() and onChange() were implemented as a channelFlow this is a somewhat problematic builder, since (quoting the docs):

A channel with the default buffer size is used.

This means that when we call send(), it doesn't necessarily wait for the downstream consumer to process that event, leading to the query potentially running more often than necessary (the behavior we want is to internally buffer upstream changes while a downstream consumer is busy, and then once it handled the previous event we'd re-run the query if there were any changes in the meantime).
I've refactored the implementation to use a regular flow, ensuring that we won't start running a second query until results of the first one are handled. I've also simplified the implementation to reuse more logic between watch and onChange.

Additionally, this updates some attachment tests:

• The testSkipFailedDownload test asserts a visible transition from QUEUED_DOWNLOAD to ARCHIVED. Depending on how often the SELECT * FROM attachments flow emits items though, it's possible that the QUEUED_DOWNLOAD state is not observable. This updates the test to reflect that.
• Because the actual attachment sync and the observed SELECT * FROM atttachments flow are decoupled and asynchronous, it's possible that a write to the todos table schedules a query that, at the time it runs, sees a different write. This has been causing a few failures as well, a distinctUntilChanged() deflakes those tests.
• I also noticed some failures related to waitFor timing out, so this adds more logs around that and replaces the waitFor calls in attachment tests.