Hardening lifecycle-state-store, name-store, and oci-hooks

[apostasie]

First commit

Although #3350 got closed as a fluke, the error message seen in the logs clearly shows that the lifecycle state file could end-up messed-up, triggering a failure.

This does two things:

• write atomically, so that we never end-up with half-written / hosed JSON in the file
• tolerate load errors: this file is not critical and purely informative and should soft fail with a warning instead

This will make the LifecycleState more robust and resistant to host shutdowns.

Note: we should consider a review of all our filesystem based stores and consider moving to atomic writing for them.

Second commit

Fix #3351 by making us resistant to the condition and allowing a container to acquire a "broken" name.

This obviously should never happen, but it does ^.

Third commit

Fix #3357 - does not fix the underlying issue, which is possibly upstream - but effectively mitigate it.
Since onPostStop is (wrongly?) called after an error in onCreateRuntime - while the container will NOT be deleted - we inhibit the normal cleanup flow in that case.

[apostasie]

Failure is the usual IPFS/compose.

[apostasie]

The Kube test (and underlying condition) is clearly racy.
This test is getting annoying (especially since right now we are just "confirming that a certain bug exists").
We might want to disable it for now (or fix the issue...). cc @lingdie

[apostasie]

Grumble grumble.
Interactive rebase...
Should be ready in a second.

[apostasie]

@AkihiroSuda CI is green. Mistake addressed.

[AkihiroSuda]

Thanks