fix(azure): single-chunk text stream misclassified as tool-call response

[alvinttang]

Summary

• Bug: AzureAIChatCompletionClient.create_stream() uses len(content_deltas) > 1 to distinguish text responses from tool-call responses. When a model returns a short answer in a single streaming chunk, content_deltas has exactly 1 element, the > 1 check fails, and the code falls through to return an empty tool-call list [] as content — silently discarding the actual text (or misassigning it to thought).
• Fix: Rewrite the content/tool-call branching to match the OpenAI client's correct logic: check if full_tool_calls: first, else treat as text content.
• Test: Add regression test that streams a single-chunk text response and asserts it is returned as string content in the CreateResult.

Root cause

# BEFORE (wrong) — line 562 of _azure_ai_client.py
if len(content_deltas) > 1:        # <-- fails when exactly 1 chunk
    content = "".join(content_deltas)
else:
    content = list(full_tool_calls.values())  # <-- returns [] for text responses

# AFTER (correct) — matches OpenAI client pattern
if full_tool_calls:
    content = list(full_tool_calls.values())
else:
    content = "".join(content_deltas) if content_deltas else ""

Test plan

• New test test_azure_ai_chat_completion_client_create_stream_single_chunk — mocks a single-chunk text stream, asserts CreateResult.content is a string (not an empty list)
• Existing Azure AI model client tests continue to pass
• Manual verification with a real Azure endpoint returning short responses