[Docs] Add a Nemotron-CLIMB data-curation recipe #2119

Context

Child of #2118. #1727 added the end-to-end Nemotron-CLIMB recipe and #2072 refined it, but the content exists only in tutorials/text/nemotron-climb-data-curation/README.md and scripts; it is not discoverable in the Fern docs.

Scope

Document the complete workflow: embedding, clustering, pruning, tokenization, mixture construction, proxy-model training/evaluation, and prediction.

Source material

• Nemotron-CLIMB data curation recipe #1727
• Minor updates to CLIMB tutorial #2072
• tutorials/text/nemotron-climb-data-curation/README.md
• Scripts 1_embed.py through 8_predict.py and e2e.sh

Requirements

• State prerequisites, GPU/storage expectations, external model assets, and required Hugging Face resources.
• Explain intermediate artifacts and restart boundaries for each numbered stage.
• Document MegatronTokenizerWriter output merging and how the result is consumed by Megatron-LM.
• Separate a quick path from full-scale/production guidance.
• Link relevant embedding, semantic-deduplication/clustering, and export references.

Acceptance criteria

• A Fern page or tightly linked page set is added under the text tutorial/recipe hierarchy
• Every numbered script has inputs, outputs, key configuration, and a runnable command
• Failure/restart behavior and expensive steps are called out
• The page is added to navigation and linked from appropriate text-curation landing pages
• Commands are validated against current script arguments
• Fern checks pass

Supplemental code-freeze attribution

#1427 introduced the reusable merge_file_prefixes utility used to combine partition-level .bin and .idx output from MegatronTokenizerWriter. Include it in the tokenizer-output merging guidance and validate the documented artifact contract.

Related PRs: #1427, #1727, #2072