Whitespace is not preserved during collation

[Gregor Middell]

The current default tokenizer is greedily consuming whitespace. Instead of consuming it at the tokenizer level, whitespace should be preserved in the token and only stripped during token normalization.