There's no reason to make clients supply a PointerRounter and a
GestureArena when constructing gesture recognizers. These objects are
statics and the gesture recognizers can just grab them directly.

Also, remove the callback constructor arguments. Almost no code used
them. Instead, people seem to prefer using the `..` operator to set
callbacks on the recognizers. Removing the arguments removes a bunch of
boilerplate.

This allows us to adjust exactly which gestures we're listening for
during layout, which I'll use to kill a SizeObserver.

We were using an Offset, which represented pixels/second, but it wasn't
clear to clients whether that was pixels/ms. Now we use a Velocity class
that is explict about the units.

Fixes #1510
Fixes #785

...even after winning the arena.

This lets it cooperate with other gestures like tap.

The way I implemented this was to refactor the entire Draggable gesture
logic to use a new kind of gesture detector called
MultiDragGestureRecognizer. It works a bit like
MultiTapGestureRecognizer but for drags.

Also some tweaks to the velocity estimator.