Add redirect_stdout_to_stderr() utility for native libraries

[fly1d]

Summary

Fixes #196.

Some native model backends write log messages straight to file descriptor 1, bypassing both sys.stdout and contextlib.redirect_stdout. This is visible with the llm-llama-cpp and llm-gpt4all plugins, where llama.cpp/gpt4all emit lines such as llama_new_context_with_model: ... directly to stdout.

This adds a small reusable redirect_stdout_to_stderr() context manager to llm.utils that redirects OS-level stdout (fd 1) to stderr for the duration of the block. Plugins can wrap their native init and generate calls with it, so that noise stays on stderr and stdout keeps only the model output.

Verification

• pytest tests/test_utils.py -q: 86 passed
• ruff check llm/utils.py tests/test_utils.py: passed
• black --check llm/utils.py tests/test_utils.py: passed

The new test uses os.pipe() to observe fd-level writes directly, so it verifies the redirect works for writes that bypass Python's sys.stdout.

Prepared with OpenAI Codex (GPT-5) assistance.