More efficient implementation of Nodeset's document order

[shepmaster]

Right now, we recompute the document order each time, but that's a known inefficiency.

@Boddlnagg points out in #48:

XSLT needs to access node-sets in document order regularly, so node-sets should always be stored in that order (instead of calculating the first node in order on demand). The difficult part is to merge two node-sets when the union operator is used. That operation has to ensure that the resulting node-set is also in document order and contains no duplicates (related to #14). Maybe some inspiration can be taken from how Gecko does this.