The call overhead from separate calls typically ends up being more expensive
than just updating a few more constants.

This gives a minor performance improvement. For example, in 3DMark03 Game Test
1 this gives about a 4% improvement, in the Counter Strike: Source stress test
it's about 1%. (NVIDIA GF9600M, Intel T9550, 1024x768).

In principle the GL driver also checks the location, but it turns out to be
slightly faster to do it ourselves.

kernel32/tests: Make test_waittxempty() use overlapped IO to avoid waiting forever when a kernel driver doesn't support TIOCGICOUNT/TIOCSERGETLSR ioctl.

Most of Linux USB-serial drivers don't support these ioctls.