Expose quiet logging for embedded GGML/Metal inference

[andimarafioti]

Problem

qwentts-cpp-python is very noisy when embedded through its Python API, particularly with the macOS Metal wheel. With qwentts-cpp-python==0.4.1 and faster-qwen3-tts==0.5.2, loading a CustomVoice GGUF pair and synthesizing one short sentence emits hundreds of native lines to stderr: ggml_metal_device_init, Metal kernel compilation, [Prompt], [Pipeline], [Sample], and [Perf].

The faster-qwen3-tts CLI is quieter because it temporarily redirects the process-wide stderr file descriptor unless --verbose is set. Applications that import the Python API do not go through that CLI path. Redirecting stderr inside a long-running, multithreaded application risks hiding unrelated diagnostics.

What I checked

• The existing QwenLibrary.set_log_callback() can intercept qwentts.cpp messages, but it does not cover GGML/Metal output. In a local Apple M3 Pro probe, registering it before model loading captured 18 qwentts messages while 146 GGML/Metal lines still reached stderr.
• qt_log_set and GGML's ggml_log_set are separate process-wide native logging interfaces. The bundled macOS libggml-base.dylib exports ggml_log_set.
• The callback lifecycle should remain safe when multiple QwenTTS contexts are created in one process, as covered by the resolved Segfault in qt_init() when a log callback (set_log_callback/qt_log_set) is registered across more than one QwenTTS context #17.

Request

Expose a supported quiet/log-level option through the Python API that can be configured before native model initialization. When set to warnings/errors, it should filter both qwentts.cpp and GGML/Metal routine output while preserving warnings and errors. Keep a verbose/debug setting for diagnostics, and document the process-wide scope of the native callbacks.

Ideally faster-qwen3-tts could then use this API instead of redirecting stderr, and embedding applications could select the same behavior without implementing native logging themselves.

Acceptance checks

• Loading a Metal GGUF pair and generating speech in quiet mode does not print routine [Pipeline], [Perf], or ggml_metal_library_compile_pipeline lines.
• Warnings and errors remain visible through the configured callback/logging path.
• Verbose mode still exposes diagnostics.
• Creating and closing more than one context with logging configured remains safe (regression coverage for Segfault in qt_init() when a log callback (set_log_callback/qt_log_set) is registered across more than one QwenTTS context #17).