fix(embedder): pin encoding_format=float on OpenAI embedding calls

[crajah]

Problem

OpenAIEmbedder calls embeddings.create() without specifying encoding_format. Left unset, the OpenAI SDK negotiates base64 on its own — and gateways fronting non-OpenAI providers reject the parameter outright:

litellm.UnsupportedParamsError: vertex_ai does not support parameters:
{'encoding_format': 'base64'}, for model=gemini-embedding-001

Every embedding call fails, so ingestion cannot proceed at all behind such a gateway. Hit in practice with LiteLLM proxying to Vertex AI for gemini-embedding-001.

Fix

Pin encoding_format='float' on both create() and create_batch(). It is the API default and the portable choice, so this is a no-op against OpenAI itself while unblocking OpenAI-compatible gateways.

Notes

• Split out of feat: add PostGraph driver — PostgreSQL-backed graph backend #1777 deliberately: that PR adds a driver, whereas this touches shared core code affecting every backend, and should be reviewable — or droppable — on its own.
• Leaving it to callers is not reachable today: OpenAIEmbedderConfig exposes no way to pass the parameter through.

[linhongyu510]

Verified the current head. Both create() and create_batch() now pass encoding_format="float" explicitly, while preserving model/input handling and embedding-dimension truncation. This matches the OpenAI embeddings API contract and avoids SDK-side base64 negotiation for compatible gateways. Local validation: pytest tests/embedder/test_openai.py -q — 2 passed; an additional focused mock check confirmed the keyword on both call paths; Ruff check/format and git diff --check passed. No blocking issue found. Non-blocking: the existing tests do not assert encoding_format, so adding those assertions would make the regression coverage durable.