A number of bugs relating to guest instance lifecycle management have
been observed. These include:

- Instances getting "stuck" in a transient state, such as `Starting` or
`Stopping`, with no way to forcibly terminate them (#4004)
- Race conditions between instances starting and receiving state
updates, which cause provisioning counters to underflow (#5042)
- Instances entering and exiting the `Failed` state when nothing is
actually wrong with them, potentially leaking virtual resources (#4226)

These typically require support intervention to resolve.

Broadly , these issues exist because the control plane's current
mechanisms for understanding and managing an instance's lifecycle state
machine are "kind of a mess". In particular:

- **(Conceptual) ownership of the CRDB `instance` record is currently
split between Nexus and sled-agent(s).** Although Nexus is the only
entity that actually reads or writes to the database, the instance's
runtime state is also modified by the sled-agents that manage its active
Propolis (and, if it's migrating, it's target Propolis), and written to
CRDB on their behalf by Nexus. This means that there are multiple copies
of the instance's state in different places at the same time, which can
potentially get out of sync. When an instance is migrating, its state is
updated by two different sled-agents, and they may potentially generate
state updates that conflict with each other. And, splitting the
responsibility between Nexus and sled-agent makes the code more complex
and harder to understand: there is no one place where all instance state
machine transitions are performed.
- **Nexus doesn't ensure that instance state updates are processed
reliably.** Instance state transitions triggered by user actions, such
as `instance-start` and `instance-delete`, are performed by distributed
sagas, ensuring that they run to completion even if the Nexus instance
executing them comes to an untimely end. This is *not* the case for
operations that result from instance state transitions reported by
sled-agents, which just happen in the HTTP APIs for reporting instance
states. If the Nexus processing such a transition crashes, gets network
partition'd, or encountering a transient error, the instance is left in
an incomplete state and the remainder of the operation will not be
performed.

This branch rewrites much of the control plane's instance state
management subsystem to resolve these issues. At a high level, it makes
the following high-level changes:

- **Nexus is now the sole owner of the `instance` record.** Sled-agents
no longer have their own copies of an instance's `InstanceRuntimeState`,
and do not generate changes to that state when reporting instance
observations to Nexus. Instead, the sled-agent only publishes updates to
the `vmm` and `migration` records (which are never modified by Nexus
directly) and Nexus is the only entity responsible for determining how
an instance's state should change in response to a VMM or migration
state update.
- **When an instance has an active VMM, its effective external state is
determined primarily by the active `vmm` record**, so that fewer state
transitions *require* changes to the `instance` record. PR #5854 laid
the ground work for this change, but it's relevant here as well.
- **All updates to an `instance` record (and resources conceptually
owned by that instance) are performed by a distributed saga.** I've
introduced a new `instance-update` saga, which is responsible for
performing all changes to the `instance` record, virtual provisioning
resources, and instance network config that are performed as part of a
state transition. Moving this to a saga helps us to ensure that these
operations are always run to completion, even in the event of a sudden
Nexus death.
- **Consistency of instance state changes is ensured by distributed
locking.** State changes may be published by multiple sled-agents to
different Nexus replicas. If one Nexus replica is processing a state
change received from a sled-agent, and then the instance's state changes
again, and the sled-agent publishes that state change to a *different*
Nexus...lots of bad things can happen, since the second state change may
be performed from the previous initial state, when it *should* have a
"happens-after" relationship with the other state transition. And, some
operations may contradict each other when performed concurrently.

To prevent these race conditions, this PR has the dubious honor of using
the first _distributed lock_ in the Oxide control plane, the "instance
updater lock". I introduced the locking primitives in PR #5831 --- see
that branch for more discussion of locking.
- **Background tasks are added to prevent missed updates**. To ensure we
cannot accidentally miss an instance update even if a Nexus dies, hits a
network partition, or just chooses to eat the state update accidentally,
we add a new `instance-updater` background task, which queries the
database for instances that are in states that require an update saga
without such a saga running, and starts the requisite sagas.

Currently, the instance update saga runs in the following cases:

- An instance's active VMM transitions to `Destroyed`, in which case the
instance's virtual resources are cleaned up and the active VMM is
unlinked.
- Either side of an instance's live migration reports that the migration
has completed successfully.
- Either side of an instance's live migration reports that the migration
has failed.

The inner workings of the instance-update saga itself is fairly complex,
and has some kind of interesting idiosyncrasies relative to the existing
sagas. I've written up a [lengthy comment] that provides an overview of
the theory behind the design of the saga and its principles of
operation, so I won't reproduce that in this commit message.

[lengthy comment]:
https://github.com/oxidecomputer/omicron/blob/357f29c8b532fef5d05ed8cbfa1e64a07e0953a5/nexus/src/app/sagas/instance_update/mod.rs#L5-L254