Fix bug with Azure Async filter events processing

[kapis]

• I understand that this repository is auto-generated and my pull request may not be merged

Changes being requested

When using (Async)ChatCompletionStream, add a weak validation before the accumulation, to ignore events not adhering to the ChatCompletionChunk schema.

This prevents runtime errors when the SSEs are not 100% compatible with OpenAI spec (e.g. when using Azure OpenAI with Asynchronous Filter feature).

Before this fix, the following code will result in an error:

from openai import AsyncAzureOpenAI

client = AsyncAzureOpenAI(
    api_version="2024-08-01-preview",
    azure_endpoint="<azure openai endpoint !!! with Asynchronous Filter enabled !!!>",
    api_key="<api key>",
)

async with client.beta.chat.completions.stream(
    model="deployment-name", 
    messages=[
        {
            "role": "user",
            "content": "hi",
        },
    ],
) as stream:
    async for event in stream:
        print(event)

# ...
# ContentDoneEvent(...)
# then Exception:

File ~/<path>/openai/lib/streaming/chat/_completions.py:397, in ChatCompletionStreamState._accumulate_chunk(self, chunk)
    369 choice_snapshot = completion_snapshot.choices[choice.index]
    370 previous_tool_calls = choice_snapshot.message.tool_calls or []
    372 choice_snapshot.message = cast(
    373     ParsedChatCompletionMessageSnapshot,
    374     construct_type(
    375         type_=ParsedChatCompletionMessageSnapshot,
    376         value=accumulate_delta(
    377             cast(
    378                 "dict[object, object]",
    379                 model_dump(
    380                     choice_snapshot.message,
    381                     # we don't want to serialise / deserialise our custom properties
    382                     # as they won't appear in the delta and we don't want to have to
    383                     # continuosly reparse the content
    384                     exclude=cast(
    385                         # cast required as mypy isn't smart enough to infer `True` here to `Literal[True]`
    386                         IncEx,
    387                         {
    388                             "parsed": True,
    389                             "tool_calls": {
    390                                 idx: {"function": {"parsed_arguments": True}}
    391                                 for idx, _ in enumerate(choice_snapshot.message.tool_calls or [])
    392                             },
    393                         },
    394                     ),
    395                 ),
    396             ),
--> 397             cast("dict[object, object]", choice.delta.to_dict()),
    398         ),
    399     ),
    400 )
    402 # ensure tools that have already been parsed are added back into the newly
    403 # constructed message snapshot
    404 for tool_index, prev_tool in enumerate(previous_tool_calls):

AttributeError: 'NoneType' object has no attribute 'to_dict'

Additional context & links

The primary purpose of this PR is to make this SDK usable with Azure OpenAI Asynchronous Filter feature.

E.g. Langchain internally relies on client.beta.chat.completions.stream in some cases.

An issue describing non-compatible events when using an Asynchronous Feature:

• Typing: when stream is completed, delta in ChatCompletionChunk from azure openai is None; should be ChoiceDelta  #1677

Fixes to the same problem in other popular SDK's:

• fix(weave): handle noncompliant chunk response from Azure wandb/weave#3922
• fix: skip processing of choice.delta when it is None run-llama/llama_index#16636

Other possibly related issues:

• LLM error occurred on Azure GPT-4o langgenius/dify#5790
• 'NoneType' object has no attribute 'tool_calls' TheR1D/shell_gpt#662

CC @kristapratico