Implement basic support for unwinding longjmp across managed frames to match x64 and old EH behavior.

Like other Windows platforms we get our exception handler called with STATUS_LONGJMP exception. However, unlike 64-bit platforms we don't get the original jmp_buf and return value in the ExceptionInformation parameters. In order to handle the longjmp we need to return back from the ProcessCLRException call which is currently not done anywhere else.

The solution is a bit of an abuse. We create our own setjmp trampoline from ProcessCLRException and use that to return from last CallCatchFunclet at the end of DispatchEx. In order not to interfere with C++ unwinding we need to extract the setjmp into its own function with no C++ destructors on the frame.

Also fix #115701 by calling GCX_PREEMP_NO_DTOR in PropagateLongJmpThroughNativeFrames while I am at it.

Contributes to #113985