[BACKEND-1] No backend-selection rubric; auto-selection behavior is opaque

## Problem

The documentation explains that auto-selection between the DTensor and Megatron backends exists but does not provide a general rubric for when each will be chosen. Users have no intuition for which backend will run for their workload and cannot make informed decisions about overriding the default.

## Suggested fix

Add a backend-selection rubric to the training backend documentation. Include guidance such as model size thresholds, tensor parallelism requirements, and use cases where each backend is preferred.