chore[notask]: register diar_streaming_sortformer_4spk-v2.1 GGUFs

[ishanvohra2]

🎯 What problem does this PR solve?

• The streaming variant of NVIDIA's Sortformer diarizer (nvidia/diar_streaming_sortformer_4spk-v2.1) is not yet exposed through the qvac model registry, so consumers of @qvac/transcription-parakeet can't pull it via the registry.
• Only the offline v1 GGUF and an external ONNX v2 entry exist today.

📝 How does it solve it?

• Adds three new entries to packages/registry-server/data/models.prod.json for nvidia/diar_streaming_sortformer_4spk-v2.1, one per quantization tier converted with packages/transcription-parakeet/scripts/convert-nemo-to-gguf.py:
  • f16 → ~251 MiB
  • q8_0 → ~134 MiB
  • q4_0 → ~72 MiB
• All three use engine: "@qvac/transcription-parakeet", point at s3:///qvac_models_compiled/ggml/parakeet/2026-05-20/diar_streaming_sortformer_4spk-v2.1.<quant>.gguf, link back to the HF model card, and tag transcription / parakeet / sortformer / streaming / 4spk / ggml / en.
• licenseId set to nvidia-open-model-license (v2.1 switched from the v1 CC-BY-4.0 license per the HF model card).

🧪 How was it tested?

• Ran scripts/convert-nemo-to-gguf.py against the upstream .nemo checkpoint for each quant; converter auto-detected model_type=sortformer and reported num_spks=4, tf_layers=18, tf_d_model=192, layers=17, use_bias=True for all three outputs.
• JSON validated locally (python -c "import json; json.load(open('packages/registry-server/data/models.prod.json'))").
• GGUF artifacts still need to be uploaded to the listed S3 path before the registry entries resolve.

[GustavoA1604]

/review

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (2/1)



---
*This comment is automatically updated when reviews change.*