Expose the internal embeddings endpoint as directly callable by clients

[virajmehta]

[GabrielBianconi]

• The endpoint should match a subset of the OpenAI embeddings endpoint. To start, we'll support input (string or array of strings) and model. We'll later add other input fields.
• The endpoint should be exposed both on /openai/v1/embeddings. We'll design /embeddings later.
• There will be no observability for now.
• You should be able to configure fallbacks (routing), timeouts (timeout_ms at the model or model provider level), and credential management features otherwise available for chat completion.
• We should support OpenAI (& anything else OpenAI-compatible) + Azure OpenAI Service to start.

[GabrielBianconi]

We merged an OpenAI-compatible embeddings endpoint that supports OpenAI and OpenAI-compatible model providers. It'll be available in the upcoming release (2025.8.1). We're now working on an Azure OpenAI Service model provider for embeddings.