Replace fprintf with qt_log across all library sources

[Xydane]

Replace fprintf with qt_log across all library sources

What changed

All fprintf(stderr, ...) call sites in the src/ library have been replaced
with qt_log(QT_LOG_xxx, ...) across 27 header files.

Headers that did not yet include qt-error.h gained the include; the
now-redundant #include <cstdio> was removed from each of those files.

Why it was needed

Previously, wrapping the library in a higher-level binding meant that load
progress, warnings, and fatal errors would bypass whatever logging framework
the host application used and land on stderr unconditionally. With this
change, a single qt_log_set(cb, user_data) call redirects everything --
Python logging, Rust tracing, Android logcat, or any other sink.

Severity mapping

Level	Used for
QT_LOG_ERROR	FATAL / OOM / file I/O / graph compute failures
QT_LOG_WARN	Recoverable surprises (missing optional GGUF keys, unknown tokens, unsupported-but-fallback paths)
QT_LOG_INFO	Normal load and synthesis cadence (model loaded, N layers, weights MB, samples written, etc.)
QT_LOG_DEBUG	Tensor dump shape lines and GGML dedup notices

Intentional exclusions

• The single std::fprintf inside qt_log's own stderr fallback path (qwen.cpp) is untouched -- it is the fallback implementation itself.
• The tools/ CLI executables are out of scope; their fprintf output is direct terminal UX, not library diagnostics.