docs: add self-contained multi-LoRA RL example (SGLang + TensorRT-LLM backends)

[kiddyboots216]

What

Adds a single self-contained example — examples/server/multilora/multilora_rl_recipe.py — for RL-style training of many LoRA adapters over one shared base model, with rollouts served by a pluggable inference engine.

The flow: create N LoRA adapters (heterogeneous rank/optimizer) → train each a few steps → export → make servable → sample per-adapter rollouts (with per-token logprobs). It has no sibling imports and inlines its session specs, so it reads and runs as one file.

Two rollout backends implement the same RolloutEngine interface:

• SGLangEngine — HTTP: the training server's create_sampling_session + SGLang /generate with per-request lora_path + return_logprob.
• TRTLLMEngine — TensorRT-LLM's in-process LLM API + LoRARequest + SamplingParams (moe_backend="CUTLASS", per-expert PEFT adapters).

The module header is a runbook: how to launch the XORL training server (using examples/server/configs/lora/qwen3_30b_a3b_lora.yaml), launch the rollout engine, and run the recipe.

Notes

• Example only — no library code changes.
• Lints clean under the repo's pre-commit hooks (ruff, ruff-format, codespell).