twice termination delay only after SIGALRM, otherwise many separate SIGCHLD
and SIGIO signals quickly increase delay to the level when SIGKILL is sent

the flags should be bit-wide only

r2909, r2910, r2922, r2923, r2924, r2925, r2929:

various win32 fixes:

*) use no-threads for Unix builds only
*) Win32 returns ERROR_PATH_NOT_FOUND instead of ERROR_FILE_NOT_FOUND
*) add trailing zero to a file name in ngx_win32_rename_file()
*) fix logging in ngx_win32_rename_file()
*) allow shared memory segments more than 4G
*) fix memory leak in successful case
*) log shared memory name in failure case
*) test that zone has the same addresses in different processes
*) add drive letter for Win32 root path
*) log GetExitCodeProcess()'s errno
*) test premature process termination
*) fix debug logging
*) exit if no workers could not be started
*) do not quit old workers if no new workers could not be started
*) a signaller process should stop configuration processing just after
   it is able to get pid file, this allows to not open log files, etc.
*) win32 master process had aready closed listening sockets

fix XSLT filter in SSI subrequests

initialize use_event field in open file cache
the bug had been introduced in r2071

fix a pool growing and CPU usage for multi-requests keepalive connections:

*) keepalive_requests
*) try to reuse last 4 free large allocation links
*) do not test a pool block space if we can not allocated from
   the block 4 times