Fix #1400
Fix #1382

When the docker runner starts a test run it does a few things:

1. Wait for the instance containers to complete,
2. Setup a "pretty printer" that forwards instance's stdout, BUT ALSO tries to decode events from stdout and update the outcomes,
3. Listen to the events sent by test instances and update the outcomes on success.

When the containers exits (1) testground stops everything related to collecting the outcomes (3).
This is fine-ish with the go-sdk because it dumps events to stdout as well, so the outcomes will be read thanks to (2).

Logging events on stdout and trying to decode the JSON looked like a legacy hack: this was not implemented in the rust-sdk. See my notes on the pretty printer. The rust-sdk triggers a race condition:

1. test instance sends "success event" to the sync service
2. test instance exits
3. Testgound shuts down the test (and the outcome collection part)
4. Testground receives the "success event", but does not listen to it anymore.

I believe we've seen this error on go-sdk, less often, see this comment.

This PR rework the local docker runner so that a test is complete if and only if:

• All the instances have exited AND all the outcomes have been received
• All the instances have exited AND a timeout of x seconds have been reached (prevents hanging when an instance crashes)
• OR the test timeout has been reached.

See runs:

• rust-sdk examples with regular testground: https://github.com/laurentsenta/testground-stability/actions/runs/2747857836
• rust-sdk examples with this patch: https://github.com/laurentsenta/testground-stability/actions/runs/2747849325

Todo

• Spike the patch,
• Verify that the patch doesn't break other components,
• Go through all other TODOs (dealing with errors, etc),
• Double check the outcome outputs and the integration tests results,
• Look into the testing situation.