Domain Distribution Analysis + Automatic Rebalancin

What: Add a pipeline stage that (1) computes domain distribution across a dataset using an existing domain classifier, and (2) resamples documents to hit a target domain mix via upsampling/downsampling.
Why:
All carefully control domain ratios (code, math, science, web, books, etc.). Random sampling from a web crawl over-indexes low-value domains. Manual rebalancing is brittle; an automated stage that reads a target distribution config and resamples accordingly is essential for reproducing frontier training mixes.

Definition of Done:

• DomainRebalancingStage under nemo_curator/stages/text/
• Accepts a YAML/dict target domain distribution (e.g., {code: 0.15, math: 0.10, web: 0.60, ...})
• Reads domain label from document metadata (output of existing DomainClassifier or LLM Classifier model)
• Supports upsample-with-repetition and downsample-by-discard strategies
• Emits before/after distribution report
• Ray-native; operates on partitioned datasets without full shuffle
• End-to-end test: input skewed corpus → verify output matches target distribution within 1%?

Start in nemo_curator/stages/text/ and trace the existing DomainClassifier or LLM Classifier output to confirm the document metadata label. Define the target YAML/dict and partitioned Ray processing behavior, including repetition, discarding, and before/after reports. Done means an end-to-end test shows a skewed corpus reaches the requested domain mix within 1% without a full shuffle.