Centralize vLLM setup utilities across inference stages

Context

The Nemotron-Parse inference stage (nemo_curator/stages/interleaved/pdf/nemotron_parse/inference.py) contains vLLM setup code (engine args, sampling params, GPU configuration) that is useful for any inference stage using vLLM.

Raised in PR #1693 (VibhuJawa): "Do we have a common way across curator to setup vLLM? If not, we should make this a common place/util to start. Given this is probably the scale at which we have tested this the most GPUs, I don't think we should duplicate this logic."

Proposed work

• Extract vLLM engine initialization, sampling parameter construction, and GPU/TP configuration into a shared util (e.g. nemo_curator/backends/vllm_utils.py or similar)
• Refactor NemotronParseInferenceStage to use the shared util
• Document as the canonical starting point for future vLLM-based inference stages

Start in nemo_curator/stages/interleaved/pdf/nemotron_parse/inference.py and inspect how NemotronParseInferenceStage builds vLLM engine arguments, sampling parameters, and GPU or tensor-parallel configuration. Extract the reusable setup into a shared backend utility, refactor this stage to use it, and document that utility as the canonical starting point for future vLLM inference stages.