Classifier perf needs improvement

[KirillOsenkov]

The classifier is now naive and parses the whole document on each keystroke, then visits the entire tree just to classify a single line. We should consider doing less work somehow, either by using TextMate or just the scanner if at all possible.