Track: MLOps framework starter templates

[ulises-jeremias]

Problem

The proposal expands MLOps into 12 templates: 9 data/framework combinations plus
3 distributed variants. That would create a large maintenance burden and many nearly
identical files.

CPA should use templates for framework-level project topology and extensions for data
modalities, distributed execution, CI/CT/CD, and local infrastructure.

Solution

Create an MLOps template track with a small number of base templates.

Recommended order:

1. Design the common MLOps contract first.
2. Implement mlops-sklearn-starter as the first CPU-friendly vertical slice.
3. Add mlops-pytorch-starter after the common contract is proven.
4. Add mlops-tensorflow-starter after the common contract is proven.

Track constraints:

• New templates must meet the CPA template quality bar.
• Default tests must be CPU-only and fast.
• No mandatory GPU/CUDA dependencies.
• No external datasets, network calls, or remote MLflow server in tests.
• Secrets must be placeholders only.
• GitHub Actions for CI/CT/CD must be delivered by an extension, not embedded in the
  base templates.

Each framework template must include a coherent project architecture, not just a
single training script.

[ulises-jeremias]

This item is part of the MLOps roadmap. If you'd like to work on it, comment here.