fix: include input_tokens in message_delta events for Anthropic compatibility

[dracpet]

Problem

Some Anthropic-compatible clients (e.g., AgentScope/QwenPaw) expect input_tokens to be present in the message_delta event's usage object. Currently, stream_from_openai_result only includes output_tokens, which causes TypeError: '>' not supported between instances of 'NoneType' and 'int' in frameworks that parse usage.input_tokens from the final delta event.

Solution

Add input_tokens to the usage dict in the message_delta event, matching the value from message_start.

Changed

# Before
"usage": {"output_tokens": usage.get("completion_tokens", 0)},

# After  
"usage": {
    "input_tokens": usage.get("prompt_tokens", 0),
    "output_tokens": usage.get("completion_tokens", 0),
},

Verification

Tested with:

• ✅ Simple streaming & non-streaming
• ✅ Tool definitions & tool results
• ✅ Multi-turn conversations (3 turns)
• ✅ Large system prompts (skill injection simulation)
• ✅ Minimal responses (edge case)
• ✅ Stop sequences
• ✅ Concurrent requests (3 simultaneous)
• ✅ AgentScope/QwenPaw integration

All 10 test scenarios pass without regression.