[Docs] Document Nemotron-Parse inference metrics, engine tuning, and Ray fanout

[lbliii]

Context

Child of #2118. #2054 added user-configurable engine_kwargs, richer inference metrics, Ray fanout behavior, and improved vLLM startup retries to the Nemotron-Parse PDF pipeline.

Required coverage

• engine_kwargs passthrough to vllm.LLM, with safe examples such as gpu_memory_utilization and max_num_batched_tokens
• Metrics: prompt/output tokens, output characters, wall time, truncated/empty outputs, retry count, pages/s, and output-tokens/s
• How to aggregate metrics using TaskPerfUtils
• PDFPartitioningStage fanout behavior under Ray Data
• Port-collision retry behavior and what users should do after retries are exhausted

Acceptance criteria

• Nemotron-Parse concept/tutorial/API content reflects all new fields
• A tuning example and a metrics-inspection example are included
• Backend-specific behavior is clearly labeled
• Defaults are checked against current code
• Related benchmark guidance is linked without exposing internal-only infrastructure
• Fern checks pass

Related PR: #2054