kernel32/tests: WaitCommEvent may sometimes return immediately on a busy system when write operation to a COM-port has already completed.

It's not enough for some hardware configurations.

kernel32/tests: Accept the case when WriteFile is able to complete the request immediately for an overlapped COM-port write.

The behaviour depends on the driver and hardware used.

This needs to work differently from CorExitProcess because we need to wait for
foreground threads and abort the rest.

Having two instances of Mono loaded at once is problematic, and now that we
use the 2.0 runtime to support 1.x there's no need to support old versions.

ole32: If an object exposes IExternalConnection then don't take down the stub manager if we have remaining weak refs in a call that would otherwise ignore them.

ole32/tests: Add some tests for CoReleaseMarshalData with a combination of normal and weak marshals.

To make DDLOCK_DISCARDCONTENTS actually do anything. The MAPPABLE flag
is now redundant.

They are not needed any more, resource_access_from_pool takes care of
this.

It segfaults on Win7 + Nvidia Kepler and apparently returns different
return values on Windows 8. Since this functionality isn't supported on
any HW I doubt games will care.