Add distributed training MLOps extensions

[ulises-jeremias]

Problem

The proposal includes three distributed templates. In CPA, distributed execution is an
advanced capability layered onto a framework starter, not a base project topology.

Solution

Create distributed training extensions after the corresponding framework templates
exist.

Extensions:

• mlops-sklearn-distributed: Dask/Ray support for parallel preprocessing and
  hyperparameter search.
• mlops-pytorch-distributed: DDP/torchrun support with single-process fallback.
• mlops-tensorflow-distributed: tf.distribute support with local strategy fallback.

Required docs for each:

• template/docs/DISTRIBUTED_TRAINING_GUIDE.md
• template/docs/README.md.append
• local execution guide;
• cluster/GPU runner notes;
• troubleshooting for missing GPUs, process launch failures, and metric aggregation.

Constraints:

• No GPU/CUDA dependency in default CI.
• Tests must run in CPU/single-process fallback.
• Distributed launch should be config-driven, not hardcoded.
• Add incompatibleWith if two distributed extensions conflict on the same files.

Acceptance criteria:

• Each extension has deterministic tests for fallback behavior.
• Docs explain how to move from local smoke tests to real multi-node/multi-GPU runs.

[ulises-jeremias]

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