The `sytest (bullseye, multi-postgres, workers, asyncio)` CI job is flaky

[anoadragon453]

The sytest (bullseye, multi-postgres, workers, asyncio) job is extremely flaky, and often causes CI is fail. This confuses contributors and places a burden on the core developers, and leads to us not looking at the Sytest CI very closely, as we're used to it failing.

Recent examples of this job failing on random batches of tests (on develop):

• https://github.com/element-hq/synapse/actions/runs/15112541464/job/42475576539#step:8:1
• https://github.com/element-hq/synapse/actions/runs/15413705036/job/43371861048#step:8:1
• https://github.com/element-hq/synapse/actions/runs/15297243328/job/43029397796#step:8:1

I propose we either disable this configuration in CI, or continue to let it run (but ignore it if it fails). Given I don't think we have time to debug it in the near future, I'd opt for the former. But we should note that doing so may lead to us forgetting to ever enable it again.

But the current state is not sustainable.

[anoadragon453]

I've created #18506 as an example of what "continue to let it run (but ignore it if it fails)" would look like.

[MadLittleMods]

From the weekly backend meeting, the conclusion was to just give access to people on the team who need to retry and nudge CI.

The flakes aren't specific to asyncio and the asyncio version probably just changes timing to show off another cohort of flawed tests. We care about asyncio because of the compatibility with Rust. I'm not sure on which way async code is being supported between the languages; but for reference we have some adapters between Tokio and the Twisted reactor (see #18357) and https://github.com/sandhose/pyo3-twisted/ which unlocks using async Rust code in Synapse.

Ideally, we'd fix the flakes but doubtful that will happen with the current mode of operation.

Personally, if we adopted a policy of fixing flakes as we see them, I would not want to fix Sytest and would just port it to Complement for my own sanity.

One thing that would be nice is to maintain a full list of flakes somewhere. Regardless of how we fix them, we will need a list at some point.