feat(embedding): add native Google Gemini embedding driver

[minhnvc]

Summary

• Adds GeminiEmbeddingDriver using the v1beta batchEmbedContents API (not OpenAI-compatible — uses x-goog-api-key header)
• Adds gemini-embedding-001 (3072 dims) and text-embedding-004 (768 dims) to infer_dimensions
• Extracts warn_external_api() helper to deduplicate the locality-check warning that was copy-pasted between the Gemini and OpenAI-compat branches of create_embedding_driver
• Promotes batch size to module-level GEMINI_BATCH_LIMIT const with doc comment
• Hoists URL construction outside the chunk loop to avoid redundant allocations

Config

[memory]
embedding_provider = "gemini"
embedding_model = "gemini-embedding-001"
embedding_api_key_env = "GEMINI_API_KEY"

Test plan

• cargo build --workspace --lib — clean compile, 0 new warnings
• test_create_embedding_driver_gemini_requires_api_key — missing key returns MissingApiKey
• Live test: start daemon with GEMINI_API_KEY set, send agent message, verify no embedding warning in logs

🤖 Generated with Claude Code

[jaberjaber23]

Thanks @minhnvc — the Gemini embedding driver is something we want, but after comparing with #997 (which landed slightly earlier as a parallel effort) we're going to land #997 as the canonical implementation because it adds:

• batch → single embedContent fallback (Gemini's batchEmbedContents isn't available in every region/plan)
• GEMINI_API_KEY / GOOGLE_API_KEY env both accepted
• google alias alongside gemini in create_embedding_driver

Your implementation is cleaner and smaller, and the GEMINI_BATCH_LIMIT = 100 constant + x-goog-api-key header + warn_external_api() helper are all nice touches. If you'd be willing to submit a follow-up PR adding those specific pieces on top of #997 (once it lands), that would be great.

Closing in favor of #997 — not because your code is wrong, just to avoid duplicate review rounds on two nearly-identical drivers. Really appreciate the contribution.