Increasing DLL's reference count from the trampoline function makes it
prone to race conditions. The thread can start executing after we have
already returned from _beginthread[ex]() and the DLL might have been
freed.

Fixes rare crash on launch with Baldur's Gate 3.
Signed-off-by: Arkadiusz Hiler <ahiler@codeweavers.com>
Signed-off-by: Piotr Caban <piotr@codeweavers.com>
Signed-off-by: Alexandre Julliard <julliard@winehq.org>