docs: Show how to start RL from an existing SFT LoRA adapter

[philippnormann]

Summary

Add docs showing how to initialize RL from an existing SFT LoRA by passing the adapter directory as the base_model when constructing art.TrainableModel. Includes a minimal example and concise motivation.

Changes

• docs/fundamentals/art-client.mdx: add “Initializing from an existing SFT LoRA” section with TrainableModel example and “Why this?” (warm-start, small-model stability).
• docs/getting-started/faq.mdx: add FAQ entry with short TrainableModel snippet.

Motivation

Many users fine-tune with SFT (e.g., Unsloth/PEFT) and want to continue with RL; pointing base_model on TrainableModel to the adapter directory is the simplest path and improves early training, especially for small models.

[bradhilton]

LGTM, any feedback @arcticfly?

[arcticfly]

Looks good! Thanks :)