Streamline internal sleep and call_later

[almarklein]

Relates to #146

This simplifies the code paths related to async sleep and call_later a bit. In particular:

• Backends use normal timers for call_later. Yes this is imprecise on Windows, but that's expected.
• The rendercanvas.asyncs.sleep() always uses a normal sleep. Yes, imprecise on Windows, but expected.
• Add rendercanvas.asyncs.precise_sleep() for where we need precise sleep: one instance in the scheduler.
  • On not-windows, this does the same as sleep()
  • On windows, this uses call_later_in_thread in combination with call_soon_threadsafe.

Why? I was looking at the code paths related to sleep, the call_later_in_thread, and the _rc_call_later in the backends. I found the code paths pretty confusion, also which code takes care of having a precise sleep. In fact, the qt precise timer was never actually used, because for that to work, asyncs.sleep would have to detect that the qt loop is active and then not use the threaded timer but the asyncadapter's sleep. Doing that felt like it would become even more complex.

This PR also contains a few minor tweaks that I added when investigating yesterdays bug, but forgot to push to the pr.

[almarklein]

Just to be clear:

• The scheduling to reach target FPS works just like it did before (now using precise_sleep in the scheduler).
• What will be less precise now (on Windows):
  • code that uses await sleep(..) from rendercanvas.utils.asyncs
  • code that uses loop.call_later()
• Does that matter? I think not, because its expected behavior in Windows. For cases where the precision is needed, users can use the new precise_sleep.