[agents] compute-embeddings: Implement batch and async execution

[eolivelli]

Sending batches of requests to the API allows you to reduce the number of calls and mitigate the risk to fall into errors due to rate limits.

Summary:

• now the compute-embeddings agent sends API requests in batches and it is fully async
• unfortunately the OpenAI Java client is not async, so actually only Vertex AI and Hugging Face REST are really async
• introduced a new parameter batch-size with default value 10 in the compute-embeddings agent
• introduced a new parameter flush-interval with default 0 (ms) in the compute-embeddings agent
• the API call is executed on reaching batch-size pending records or flush-interval ms of wait.
• the feature is disabled by default as it could add extra latency to the chatbot pipelines

Interesting things to document:

• you should enable batching (batch-size > 0 and flush-interval > 0) on the pipelines that are not sensitive to latency (like background text processing)
• on the chatbot pipelines it is better to keep the defaults (flush-interval to 0) in order to reduce the latency

pipeline:
  - name: "compute-embeddings"
    id: "step1"
    type: "compute-ai-embeddings"
    output: "chunks-topic"
    configuration:
      model: "text-embedding-ada-002" # This needs to match the name of the model deployment, not the base model
      embeddings-field: "value.embeddings_vector"
      text: "{{% value.text }}"
      batch-size: 10
      flush-interval: 1000

Backward compatiblity
This PR is 100% backward compatible because we set flush-interval to 0 by default, so the effect is that we always compute only 1 embeddings at a time.