nexus: split out control plane zones as separate artifacts

[iliana]

The primary change here is adding a control knob to ArtifactsWithPlan::from_stream that either splits out the artifacts or combines all the zones into a single tarball.

ControlPlaneZonesMode::Composite is the current behavior and is used in Wicket and existing unit tests. For the moment this doesn't support handling repositories with artifacts already split out, but the semantics of this setting are that if/when we do that, it will combine them into a single artifact for Wicket.

ControlPlaneZonesMode::Split enables the new behavior and is used in Nexus. This makes the individual zone images available in the TUF Repo Depot and therefore immediately usable by Sled Agent when creating zones.

Additionally, we add a check to assembling a TUF repo to ensure that no individual zone images have the same hash; since the fake zones created during tests had the same hashes, we learned such a repo would not be accepted by Nexus due to this check:

omicron/update-common/src/artifacts/update_plan.rs

Lines 875 to 882 in e5bf48f

	let by_hash_slot = match self.by_hash.entry(artifact_hash_id) {
	Entry::Occupied(slot) => {
	return Err(RepositoryError::DuplicateHashEntry(
	slot.key().clone(),
	));
	}
	Entry::Vacant(slot) => slot,
	};

Closes #4411. (If we want to split out the control plane zones in actual repos we can open a new issue for that.)

[iliana]

Note to self: this doesn't set the artifact name correctly (it uses zone1.tar.gz instead of zone1). Will fix...

[jgallagher]

LGTM, just a couple nits to take or leave.

[jgallagher]

Nit - we shouldn't be including the #[source] error in the string (https://github.com/oxidecomputer/slog-error-chain?tab=readme-ov-file#aside-embedding-source-error-strings). I realize this was here before, but since we're patching up the message, can we drop the : {err} here and in ZoneVersion?

[iliana]

Maybe not right now -- the one place we use this error we're just running err.to_string() on it, and apparently thiserror's Display implementation doesn't support including the error chain with the alternate format like anyhow does.

[jgallagher]

Can we replace that with InlineErrorChain::new(&err).to_string()?

[jgallagher]

Maybe add (except Zone) like the other comments? I'm not sure how valuable that really is, but it does seem like the comment as written is now incorrect.

[jgallagher]

The unit test failure seems legit - does data need to be a tarball containing an oxide.json?