[Docs IA] Modularize large workflow pages and multimodal synthetic navigation

Parent tracker: NVIDIA-NeMo/Curator#2118

Goal

Reduce mixed-intent documentation pages and establish a clearer home for multimodal synthetic-data workflows after the 26.07 release.

Candidate improvements

• Split the Inference Server page into backend selection, Ray Serve, Dynamo aggregated, Dynamo disaggregated, and configuration/troubleshooting articles.
• Split Translation into a task guide and a TranslationStage/quality/reference article.
• Separate video caption generation/model selection from preview generation if user testing confirms the two-task split.
• Evaluate a modality-neutral synthetic or multimodal section for Nemotron OCR and future image-text generation workflows.
• Reuse the Nemotron-CLIMB overview plus nested stage-reference pattern where recipe documentation needs both orientation and exhaustive controls.

Timing

Non-blocking information-architecture debt. Address after the 26.07 documentation PRs merge and publish; do not delay the release solely for these page splits.

Acceptance criteria

• Confirm proposed splits with page analytics, support questions, or a focused content audit.
• Preserve stable URLs with redirects for any moved pages.
• Keep overview pages task-oriented and move exhaustive parameter material into reference pages.
• Update navigation, incoming links, and release-note links atomically.
• Run Fern and broken-link validation after each move.