I added overflow hidden on the html and body tags in a previous commit to prevent rubber banding when scrolling the editor. But it broke the ability to scroll specs. This fixes that.
Rather than managing the visibility directly, a class makes it easy. The cursors will always be toggling this class on and off in the background, but only on the focused editor will it actually have an effect.
Also, when stopping blinking, set the cursor's visibility to whatever it is on the underlying model. This fixes issues where the cursor would blink when there was a selection.
This is actually more efficient than the CSS animation we were doing previously, because it doesn't force the cursor to be sampled at 60 FPS for something that changes around twice a second.
This will be used by the status bar and other expensive code that needs to respond to the buffer changing, but that we don't want running on every single keystroke when the user is typing quickly.
This is the fastest way I've found to get the nextTick behavior, which allows us to aggregate multiple requests for display update (from selection changes, cursor movement, line changes) etc into a single call at the end of processing an event.
Then in the editor, we update and autoscroll only the views that need it. Autoscrolling is suppressed when resetting the display so that we honor the scroll positions of the active EditSession.
We no longer call `updateDisplay` immediately when the cursor is attached. Instead, we always attach and remove cursor views from `Editor#updateDisplay`, so we will decide whether to autoscroll at that level.
The goal is that no model operation will ever modify the DOM directly. We will always request a display update, which will happen on the next animation frame.
$base works as $self worked previously. $self will need to be fixed. If inside
a embedded grammar $self refers to the embedded grammar while $base refers to the
overall grammar.
