Add OpenAI API embedder behind optional extra

[Dawn-OuYang]

Title:
Add OpenAI API embedder behind optional extra

Body:
Closes #10

This adds the first hosted API embedder backend for Retrieval Lab. The new
api:<model> selector currently targets OpenAI embedding models and keeps the default
install keyless by importing the OpenAI SDK only when the API embedder is selected.

Changes:

• Add OpenAIEmbedder in retrieval_lab.embedding.api.
• Add the optional [api-embed] extra with the OpenAI SDK dependency.
• Wire --embed-models api:<model> and geometry --embed-model api:<model> through the CLI.
• Reuse the existing EmbeddingCache content-hash cache for hosted embeddings.
• Add mocked/keyless tests for the API embedder and missing-extra CLI error.
• Document installation, OPENAI_API_KEY, and provider availability/rate-limit caveats.
• Update CHANGELOG.md.

Duplicate check:

• Checked open PRs mentioning #10; only an unrelated Dependabot PR is open.
• Checked open PRs for api-embed, OpenAIEmbedder, and hosted embedding; no duplicate PR found.

Validation:

• python -m pytest -q
• ruff check src tests
• git diff --check
• python -m py_compile src/retrieval_lab/embedding/api.py src/retrieval_lab/embedding/__init__.py src/retrieval_lab/cli.py tests/test_embedding.py tests/test_cli.py

Note:

• ruff format --check src tests was not used as a gate here because the latest local Ruff
  wants to reformat many pre-existing files outside this patch. This PR leaves unrelated
  formatting untouched.

[AshwinUgale]

Thanks for this, @Dawn-OuYang — really clean work.

Lazy-importing the OpenAI SDK keeps the default install and tests keyless, reusing the existing content-hash cache gives hosted embeddings free dedup across a sweep, and the injected-fake-client tests cover the adapter without a real key. Docs and CHANGELOG updates appreciated too.

Merging now — thanks again! 🚀