feat(extraction): support Nemotron Parse 2.0

[ChrisJar]

Summary

• make nvidia/NVIDIA-Nemotron-Parse-2.0 the Retriever-managed local default while preserving the NemotronParseV12 compatibility alias
• restore the tied decoder output weight required by the compact Parse 2.0 checkpoint before vLLM loads it, and pin the tested Hugging Face revision
• honor explicit local nemotron_parse_model selection and document managed local, hosted Build, self-hosted NIM, and self-hosted vLLM contracts
• add regression coverage for local model selection, compact-checkpoint weight tying, and the Parse 2.0 NIM request contract

Root cause

Latest main always instantiated the v1.2 local wrapper and ignored nemotron_parse_model. Loading the compact Parse 2.0 Hugging Face checkpoint with stock vLLM also left lm_head independent from the tied decoder embeddings, producing empty generations. The official Parse 2.0 NIM uses the tagged chat contract and its 9,000-token context includes the control prompt, so Retriever must retain an 8,192 output-token request rather than requesting the entire context for output.

Validation

• 3126 passed, 118 skipped, 12 deselected across nemo_retriever/tests
• core Ruff checks (E4,E7,E9,F) pass on all changed Python files
• Retriever-managed local actor, no endpoint and no manual server: first page of all 20 jp20 PDFs, 0 errors; 14 text pages, 2 tables, 7 charts, 1 infographic
• nvcr.io/nim/nvidia/nemotron-parse-v2.0:2.0.8-variant through Retriever HTTP integration: same 20 jp20 pages and routed counts, 0 errors
• git diff --check passes

Strict MkDocs rendering was not run because MkDocs is not installed in the available environments.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.