Executables that use libbullet are now 1 MiB smaller (stripped!).

Devices were not properly reset at each frame, which caused
button_pressed() and other functions to fail.

Now we can call functions that need to access GPU resources explicitly
on the draw thread.

This lets us get rid of m_autorelease and m_destroy and we’ll be able
to implement init_game/init_draw later, for stuff that must happen
on the render thread.

We need to ensure the inter-thread messaging queues are empty (more
specifically, not locked) before destroying them. The best way to ensure
that is to explicitly wait for thread termination.

Controller::WasKeyReleasedThisFrame() is now keyboard->key_released() so
there is usually no need for a controller. They will completely disappear
when joystick axis binding has been ported to InputDevice.