[pa-devbot]

Follow-up from the Mac Intel PaddleOCR crash investigation (#1811, #1812, #1817).

When the command-line test runner dies on a fatal signal, the CI log ends right before the crash — the process's last stdio buffers are lost, and on macOS the .ips crash report doesn't land in the log. This made the #1811 crash take multiple CI runs to pin down.

What this adds (command-line test mode only, no GUI impact):

• Handlers for SIGSEGV/SIGBUS/SIGABRT/SIGILL/SIGFPE (Apple/Linux) that write a native backtrace of the crashing thread straight to fd 2 — unbuffered, and bypassing the app's std::cerr redirection into its logger — then restore the default disposition and re-raise, so the OS still writes its crash report and the exit code (139) still reflects the crash. A re-entrancy guard makes a crash inside the handler just die immediately. No-op on Windows (no portable async-signal-safe backtrace API there; OS crash reports cover it).
• setvbuf(stdout, _IOLBF, 8192) in test mode so the test lines reach the log as they happen instead of being lost in the full pipe buffer. The explicit size matters: with 0, MSVC's ucrt ends up with a zero-size stdout buffer and the first write crashes with 0xC0000409 (hit this during the investigation).

Proof it works: the #1817 A/B run used this handler to capture the backtrace that identified the Mac Intel crash root cause (a stack-use-after-return in the Qt socket tests, since fixed on main):

[PA-CRASH] Fatal signal, backtrace of the crashing thread:
0   SerialPrograms  ... crash_backtrace_handler + 168
1   libsystem_platform.dylib  _sigtramp + 29
3   SerialPrograms  std::function<void()>::~function() + 21
6   SerialPrograms  PokemonAutomation::(anonymous namespace)::Waiter::wait_for(...)
7   SerialPrograms  PokemonAutomation::(anonymous namespace)::Test_Sockets_RapidReconnect<ClientSocket_Qt, ServerSocket_Qt>::run()
8   SerialPrograms  UnitTestRunner::dispatch_test(...)