fix(autoscaling): extract load model from Route() and use as well when doing autoscale

[mudler]

This pull request adds support for storing and retrieving model load metadata to enable the reconciler to reliably replicate model loads during scale-up. It introduces new methods and fields for saving backend type and serialized model options, and refactors the model scheduling logic to use this metadata when loading models on new nodes. The changes also include updates to interfaces, implementations, and test fakes.

Model load metadata storage and retrieval:

• Added SetNodeModelLoadInfo and GetModelLoadInfo methods to the ModelRouter interface and its implementations, allowing the backend type and serialized model options to be saved and fetched for each loaded model. [1] [2]
• Extended the NodeModel struct with BackendType and ModelOptsBlob fields to persist this metadata in the database.

Model scheduling and loading refactor:

• Introduced the scheduleAndLoad helper method in SmartRouter to centralize the logic for picking a node, staging files, loading the model, and recording metadata.
• Refactored both Route and ScheduleAndLoadModel to use stored metadata for consistent model loading and replica creation, improving reliability during reconciler-driven scale-ups. [1] [2]

Test updates:

• Updated test fakes in model_router_test.go and router_test.go to implement the new interface methods, ensuring test coverage for the new logic. [1] [2]

Miscellaneous:

• Added necessary imports for fmt in test files.

Notes for Reviewers

Signed commits

• Yes, I signed my commits.