Allow scaling a ModelDeployment to zero replicas

[haarchri]

Description of your changes

Allow scaling a ModelDeployment to zero replicas

spec.replicas carried a minimum of 1, so kubectl scale --replicas=0 was rejected at admission. The docs name KEDA and the scale subresource in the same paragraph, and scale-to-zero is a main reason to put KEDA in front of an expensive GPU workload, so the floor broke an expectation the API itself set. There was also no way to park a deployment: withdrawing its endpoints while keeping the object in place meant tainting the hosting cluster.

Dropping the floor alone would misreport the parked state. With zero desired, an empty schedule reads as InsufficientCapacity 0 of 0 replicas scheduled, and a control plane with no clusters reads as
NoClusters, both leaving the XR not ready, so a deliberately parked deployment would look permanently broken.

Zero desired now takes a dedicated path before input resolution: nothing is composed so the existing ModelReplicas and ModelEndpoints are pruned, status.replicas reports 0 for the scale subresource, and ReplicasScheduled and ReplicasReady read True with a new ScaledToZero reason, the way a Deployment at zero replicas reports Available. The docs' Scaling section now states the 0 to 10 range and the parked semantics.

Fixes #438

I have:

• Read and followed Modelplane's contribution process.
• Run nix flake check (or ./nix.sh flake check) and made sure it passes.
• Added or updated tests covering any composition function changes.
• Signed off every commit with git commit -s.

[dennis-upbound]

Does parking actually release the GPUs?

The getting-started cast says minNodeCount: 1 # keep >=1; the autoscaler can't scale a GPU pool up from 0 for DRA pods. So on a pool sized that way, parking removes the pods but the floor keeps the nodes, and you keep paying for them. Set minNodeCount: 0 and the pool can empty, but then nothing brings it back for DRA pods — parking becomes one-way.

Savings are real in between (8 nodes down to 1), so this isn't an objection to the feature. But the docs say "a KEDA scaler can idle an expensive deployment and later restore it", and restore only holds while the floor is 1 or more. Worth saying which of the two it is.

[negz]

Few nits but LGTM overall.

[negz]

Does parking actually release the GPUs?

It releases them from this MD, making them available to others. It doesn't release them from the cluster. I think that's worthwhile.

Worth saying which of the two it is.

Not following what needs to be clarified.