lxd profiles aren't reliably applied on juju 2.9.27

[Adam Dyess]

During deployment of a charm integration tests, we use juju to deploy a bundle into a lxd cloud. This resulted in repeated failed tests which appeared sometime between 2.9.22 and 2.9.27.

It's been validated that the following with a 2.9.27 client will reproduce the issue

Reproduce:

```
git clone https://github.com/charmed-kubernetes/charm-containerd.git
cd charm-containerd
tox -e integration -- --controller <lxd-controller> --cloud <lxd-cloud>
```

two of the machines will block in the "allocating" state with the message
"required charm profile not yet applied to machine"

preventing the successful deployment of the charms.

here's the output from dump-db
https://paste.ubuntu.com/p/WZGYBMTFQF/

and the output from lxc profile list
https://paste.ubuntu.com/p/nwDrQtGFvS/

[Heather Lanigan (hmlanigan)]

Restarting the controller's jujud allows the profiles to be applied.

[Heather Lanigan (hmlanigan)]

Reproduces also with this bundle:
https://people.canonical.com/~heather/1966129-bundle.yaml

Timing windows in the new lxd profile watchers need to be investigated. Not seen in our CI tests. Though all are not running right now.

The db has the profile in the charm, it just never gets written to lxd nor applied to the correct machines. If the instanceMutator is restarted, the correct profiles get applied during the reconciliation.

[Heather Lanigan (hmlanigan)]

This bug was introduced when the lxd profile watcher code was moved from the model cache to a state watcher. The first notification of an instanceData doc does not indicate that the machine was provisioned. The InstanceID must change.

https://github.com/juju/juju/blob/juju-2.9.23/core/cache/machine.go#L168

[Heather Lanigan (hmlanigan)]

https://github.com/juju/juju/pull/13891