Fix pending output corruption when changing file buffers

[fhgffy]

Changing a file sink from a 64 KiB buffer to 128 KiB while log output is pending can corrupt the previously written message. The original reproducer produced the expected 29 bytes, but their contents differed. file_helper::set_buffer_size() resizes or releases storage that the open FILE still uses, and calls setvbuf after I/O.

Flush the pending output and reuse the existing non-truncating reopen path so the new buffer is configured before I/O. Assigning the same size is a no-op. After a size change, the rotating sink flushes callback output and refreshes its size estimate.

A size change now reopens the current pathname in append mode and invokes the existing close/open handlers. It uses the existing flush/reopen exception paths; callers must handle reconfiguration failures. This also avoids relying on runtime-specific behavior for repeated setvbuf calls (C11 draft, section 7.21.5.6).

Validation on Windows, MinGW GCC 15.1.0, C++11 Debug, with warnings treated as errors:

• The new pending-output regression fails on the unmodified base in both compiled and header-only builds (2 failed assertions out of 16 each).
• The original reproducer passes after the fix; focused tests pass in both builds (42 assertions in 6 test cases each), covering default/custom buffers, growth, shrinkage, reset, unchanged sizes, file events, and rotating size tracking.
• Full unit tests pass: compiled 725 assertions / 196 cases; header-only 746 assertions / 196 cases.
• CTest passes all 3 targets, including CMake static/shared source and installed-consumer checks.

Linux/macOS, MSVC, wchar configurations, C++20 std::format, and sanitizers have not been run locally.

[gabime]

Thanks @fhgffy