Support HuggingFace completion prompts for non-Llama models

[mitre88]

Summary

• Add a generic HuggingFace completion prompt path so non-Llama generation endpoints can construct prompts instead of failing in prompt construction.
• Replace the previous KeyError('model_tag') path with an actionable error for unsupported HuggingFace chat mode.
• Add focused tests covering non-Llama completion prompts and the non-Llama chat error message.

Why

Issue #221 reports an unhandled KeyError('model_tag') when trying to evaluate a HuggingFace model such as microsoft/phi-2. The prompt constructor referenced a missing model_tag config key before it could give a useful error or allow a completion-mode prompt.

Validation

• REDDIT=http://reddit.local SHOPPING=http://shopping.local SHOPPING_ADMIN=http://shopping-admin.local GITLAB=http://gitlab.local WIKIPEDIA=http://wikipedia.local MAP=http://map.local HOMEPAGE=http://homepage.local uv run --python 3.11 --with pytest==7.1.2 --with pytest-asyncio --with gymnasium --with playwright==1.32.1 --with beartype==0.12.0 --with tiktoken --with transformers==4.33.2 --with openai==0.27.0 --with aiolimiter --with nltk --with text-generation --with Pillow python -m pytest tests/test_prompt_constructor.py -q
• git diff --check

Note: I used Python 3.11 for local validation on Apple Silicon because Playwright 1.32.1 depends on greenlet 2.0.1, whose Python 3.10 macOS wheel is x86_64-only and failed to build locally on arm64.

/claim https://algora.io/PrimeIntellect-ai/bounties/U3xSkpanaVo8u1sT