fix: set a default LiteLLM request timeout

[fcsguyxcw]

Summary

• set a 180-second default timeout for all direct LiteLLM calls
• preserve existing global and per-model overrides
• verify the default is merged and reaches streaming chat completion requests

Why

Without a timeout, an accepted request that stops producing data can await indefinitely, so Agent Zero never reaches its exception-driven retry path. The pinned LiteLLM direct completion API accepts timeout for both streaming and non-streaming calls. stream_timeout is primarily handled by LiteLLM Router, which Agent Zero does not use for these requests.

Tests

• python -m pytest tests/test_stream_tool_early_stop.py (55 passed)
• focused post-assertion run for the default merge and streaming propagation tests (2 passed)
• git diff --check

Closes #1882