Add quiet native logging for embedded inference

[andimarafioti]

Summary

• Add log_level to QwenLibrary, QwenTTS, and from_pretrained, including quiet/warning and verbose/debug modes before native model initialization.
• Filter qwentts.cpp and GGML/Metal logs without redirecting process stderr, keeping callback trampolines alive across contexts.
• Preserve Python-installed GGML callbacks during first backend initialization and route the pinned native source's remaining direct diagnostics through its log levels during wheel builds.
• Document the process-wide behavior and add filtering and callback lifetime tests.

Closes #27.

Verification

• PYTHONPATH=src QWENTTS_CPP_LIBRARY=src/qwentts_cpp/lib/libqwen.dylib .venv/bin/python -m pytest tests -q — 65 passed, 1 skipped.
• Local Metal CustomVoice GGUF load and synthesis in warning mode, twice in one process — both generated 30,720 samples at 24 kHz; stderr had zero lines.
• Equivalent CPU run — both contexts synthesized successfully with zero stderr lines.
• Native GGML warning/error and qwentts.cpp error probes remained visible; debug Metal synthesis emitted pipeline, performance, and kernel diagnostics.