docs: clarify multi-GPU runtime setup

[SufianTA]

Summary

• document that the default examples use the normal TxAgent load_models path
• add guidance for users who manually initialize vLLM with tensor parallel settings
• call out the TxAgent initialization steps needed before run_multistep_agent()

Validation

• git diff --check
• python -m compileall -q src run_example.py run_txagent_app.py

Addresses #18