v0.4.1 — airgap embedders, Streamlit example, test-suite fix

Added

• Air-gapped / offline model loading — point embedding.model_name at a local directory of side-loaded weights and nothing is ever downloaded, for every provider. fastembed states the model recipe in YAML (new pooling knob + normalize / model_file_name); sentence-transformers and PyLate take the directory natively (the right choice for last-token models like Qwen3-Embedding) with HF offline mode forced; openai rejects a local path with an actionable error. Existing configs keep their exact pipeline hashes — nothing re-embeds. (#121)
• Ask-your-docs Streamlit webapp — examples/ask_your_docs_agent ships a themed chat UI (streamlit run streamlit_app.py) over the LangGraph agent: sidebar config, conversation memory with follow-up reformulation, code snippets in fenced blocks. The example is now Streamlit + notebook only. (#122)

Fixed

• Full-suite failures in the fast-plaid storage tests — the default-install no-torch test evicted torch from sys.modules without restoring it, crashing any later torch import in the same run. Suite is fully green. (#123)

Changed

• The ask-your-docs example defines its dependencies in a single requirements.txt (requirements.py removed). (#124)

Full Changelog: v0.4.0...v0.4.1