Wrap embedding API errors in EmbeddingClientError instead of letting them propagate raw

[nicoloesch]

Summary

• EmbeddingClient.embeddings() and the embedding_dim live probe now catch openai.APIError around the underlying .embeddings.create() calls and re-raise as EmbeddingClientError
  • callers get a richer, more actionable message instead of a raw SDK exception with no context.
• The message names plausible causes (oversized input, auth, rate-limit, connectivity) without asserting a specific diagnosis, since the true cause can vary by provider and isn't reliably detectable from the exception alone.
• No proactive token/length limit added
  • per-model limits vary and a hardcoded cap would be arbitrary and go stale.

Related to conversation here: AustralianCancerDataNetwork/omop-graph#29 (comment)