With the buffer / texture mapping acceleration bits working properly, we
can get many outstanding commands and, to avoid blocking the client thread
while waiting for free space, we need a larger queue. 16 MiB was always
enough in my testing (8 MiB wasn't in a few cases).

Keep a lower limit on 32-bit since we're often times address
space-starved there.

Tweaked by Zebediah Figura.