As part of #5119

Motivation:

generate_rollout_completions() is the primary helper API for rollout function
authors who need to call the inference backend mid-rollout (for multi-turn agent
loops). It currently hard-requires vLLM:

if not trainer.use_vllm:
    raise RuntimeError("Custom rollouts require vLLM ...")

Scope:

• Replace direct vLLM API calls (trainer.vllm_generation.vllm_client.generate(),
  trainer.vllm_generation.llm.generate()) with calls through the generation
  backend interface.
• Remove the if not trainer.use_vllm: raise RuntimeError(...) guard.
• Preserve the parameter-building helpers (_build_base_generation_kwargs,
  _build_colocate_sampling_params, _build_server_generation_kwargs) but make
  them backend-aware or extract backend-specific ones.
• Update function signature if needed to accept a backend-agnostic generation
  callable.

CC: @albertvillanova