Support streaming of choice-less chunks

[adubovik]

Currenly, all choice-less chunks (such as chunks which report usage and statistics fields) aren't immediately sent into a response stream, but their sending is delayed until a chunk with a choice is encountered.
These choice-less chunks are then merged together and sent along with this choice-having carrier chunk.

Since there is a chance that choice-less chunks will be received from the upstream as the very last chunks, the carrier chunk is chosen as the last choice closing chunk:

ai-dial-sdk/aidial_sdk/chat_completion/response.py

Lines 108 to 127 in 8abe579

	if isinstance(chunk, BaseChunk):
	is_last_end_choice_chunk = (
	isinstance(chunk, EndChoiceChunk)
	and chunk.choice_index == self.n - 1
	)
	is_top_level_chunk = isinstance(
	chunk,
	(
	UsageChunk,
	UsagePerModelChunk,
	DiscardedMessagesChunk,
	),
	)
	if is_last_end_choice_chunk or is_top_level_chunk:
	delayed_chunks.append(chunk)
	else:
	yield _create_chunk(chunk)

Thus, such choice-less chunks are reported at the very end, which precludes their streaming.

It's desirable to enable streaming of the statistics fields: #15
Imagine an application which calls model A, reports per-model usage for A, then fails to call model B.
Currently, the downstream won't see any per-model usage since they are reporting in the very last chunk.

A similar delaying technique is used in adapter-openai to eliminate choice-less chunks.

Q: Why do we delay sending choice-less chunks in the first place?
A: Because we didn't know how to introduce the missing choices field the best. The possible solutions are:

1. Add a fake message with an empty string content:

{
  "choices": [{"index": 0, "delta": {"content": ""}}],
  "usage": {
    "prompt_tokens": 1,
    "completion_tokens": 2,
    "total_tokens": 3
  }
}

2. Add an empty list of choices:

{
  "choices": [],
  "usage": {
    "prompt_tokens": 1,
    "completion_tokens": 2,
    "total_tokens": 3
  }
}

In the case of an empty list of choices, it wasn't clear if it could be parsed correctly downstream.
However, since then OpenAI has introduced stream_option.include_usage feature. It enables the generation of a chunk with an empty list of choices and a non-empty usage field.

This is correctly handled by popular OpenAI clients (openai and langchain libraries).

The proposal is to follow this empty list of choices convention.

TODO: ascertain that openai and langchain will be able to parse chunks with an empty list of choice and statistics field.