feat(embedding): add jina and mistral providers, fix stale install hints

[zc277584121]

Summary

• Compared MFS's embedding provider set against upstream Memsearch (the project this embeddings code was originally adapted from) and found two providers supported upstream but missing here: jina and mistral.
• jina.py: talks to Jina's embeddings REST API (https://api.jina.ai/v1/embeddings) directly over the core httpx dependency — no new extra needed. Default model jina-embeddings-v4 (2048-dim, Matryoshka-truncatable).
• mistral.py: uses the official mistralai SDK behind a new mistral extra. Default model mistral-embed (1024-dim).
• Both wired into the provider registry (_PROVIDERS/DEFAULT_MODELS/_INSTALL_HINTS), the mfs-server setup --section embedding wizard, pyproject.toml, and docs/providers.md.
• Also fixed a copy-paste leftover: openai.py, voyage.py, ollama.py, and local.py's module docstrings still told users to pip install memsearch (from when this code was adapted from Memsearch) instead of mfs-server. Swept the whole repo for other occurrences — these four were the only ones; the remaining "memsearch" mentions (README, docs/why.md, .gitignore) are legitimate references to the project itself.

Test plan

• New tests/test_embeddings_registry.py: registry lists all 8 providers with a default model each; jina raises a clear RuntimeError without JINA_API_KEY; mistral raises the registry's install-hint ImportError when the SDK isn't installed; a mocked-transport test exercises JinaEmbedding.embed() end to end (request body, auth header, response parsing).
• Verified the new mistral extra resolves cleanly (uv sync --extra mistral) and uv.lock picked it up correctly for both the mistral and all-providers extras.
• Full pytest suite passes with no regressions.
• ruff format --check / ruff check clean.