This fixes a performance problem introduced in my last commit when
max_swapchain_images == 3.
In my last commit I incorrectly stated the 3rd buffer was not being
used, after doing some additional testing I realized I was mistaken,
it can indeed be used due to the used page tracking.
Reverting that portion of my previous commit.
When max_swapchain_images was originally added it worked properly,
but was subsequently broken by using the value to specify the number
of buffers allocated.
Due to how the dispmanx driver works, only 2 buffers are ever actually
used, so the 3rd buffer in the "swapchain" ended up doing nothing.
Fix this by restoring it to the original intent, that is, if
max_swapchain_images <= 2 wait for vsync after the flip (reducing
lag), otherwise wait at the last possible moment (increasing lag).
Additionally, fix up some unnecessary void* usage where type safety
could be maintained.
It's unlikely it was causing issues, but another race existed in the
dispmanx driver, it incremented the pending page count after sending the
update request.
If a context switch happened between sending the update and incrementing
the page count, the vsync callback would decrement it from 0 and cause
the uint to overflow.
Throw in a minor C89 fix and some whitespace fixups too.
The dispmanx driver had a small race wherein if the vsync completed
between checking for a free page and waiting on the condvar, it would
hang forever waiting for a condition that would never fire.
I'm hoping this is what was causing the triple buffering lockups. In
my testing with it re-enabled and this fix, things are stable (and
much more performant than with triple buffering disabled).
