Add dimensions parameter support for embeddings endpoint (Matryoshka)

The /v3/embeddings endpoint currently doesn't support the dimensions parameter from the OpenAI API.

Many modern embedding models (like ibm-granite/granite-embedding-311m-multilingual-r2, nomic-embed-text, text-embedding-3-small) support Matryoshka Representation Learning, which allows truncating embeddings to smaller dimensions without retraining.

Request

Add support for the dimensions parameter in EmbeddingsCalculatorOV to truncate and re-normalize the output embeddings.

Example request

{
  "model": "my-embedding-model",
  "input": "Hello world",
  "dimensions": 256
}

Expected behavior

Return embeddings truncated to 256 dimensions (instead of full 768), re-normalized.

References

• OpenAI API: https://platform.openai.com/docs/api-reference/embeddings/create#embeddings-create-dimensions
• Matryoshka paper: https://arxiv.org/abs/2205.13147

Start at EmbeddingsCalculatorOV and the /v3/embeddings request handling, then trace the existing embedding calculation path. Done means a dimensions request returns truncated, re-normalized embeddings while requests without dimensions retain their full output.