fix(server): Use tokio::select when waiting for connection close/global graceful shutdown

[thesw4rm]

Solves linkerd/linkerd2#15684

BUG

Connections were not being dropped from memory after being closed. server_conn was waiting on a release shutdown signal. This was a regression likely caused by the hyper 0.14 to 1 bump in a95478a

FIX

Switch back to a tokio::select pattern. Branch predictor is optimized to check the connection being closed first, before polling for a shutdown signal.

Signed-off-by: Yadunandan Pillai contact@mail.thesw4rm.com

[raykroeker]

Thanks for sending the PR. Some feedback to consider:

1. Housekeeping. Have a look at the sign-off in CONTRIBUTING.md and DCO it'd be great to signoff on all commits.
2. Given we're no longer using graceful, drop the feature from Cargo.toml.
3. The version change to Cargo.lock isn't tied to this change.
4. Add some testing that demonstrates the validity of the tokio::select pattern.

[thesw4rm]

@raykroeker

Appreciate it!

1, Signoff passes DCO now.
2. Graceful is dropped from Cargo.toml
3. Cargo.lock looks like it is out of sync. My build keeps changing it. I've kept it 0.26.0 for now, but kubert package is 0.27.0
4. Unit tests added. They consistently fail on the old pattern, and pass on the new one. I'm not sure that this test should be kept in there though. The bug in the previous code was, from what I've understood, a one-off instance rather than behavior we expect to see again.

[thesw4rm]

Hey @raykroeker . Looks like the issue reporter confirmed that linkerd/linkerd2#15684 is likely directly related to the fix (same reproduction steps).

They're not comfortable building their own image. Do you think this PR is ready to merge, so they can confirm the fix on their end?

[thesw4rm]

@raykroeker bumping this one, if you've got the time!