Rework text parsing

[nadako]

Right now, when parsing text, we add a Text token for each character, marking whitespace ones, setting them to null and then squashing them into one Text token. This was ported from mustache.js and is needed to handle stand-alone tags on line.

I don't like it as it creates a lot of garbage and generally feels quite hacky, so I need to think of some better way to handle all this. Reworking this might accidentally fix disabled spec tests mentioned in #4.