Document should perhaps track changes directly

[Olly Betts]

Follow on from #250.

I wonder if a Document object should track the changes made, rather than copy the old data and modify it. Currently adding a new term to a document is unnecessarily O(number of terms indexing that document). The current optimisation is still useful in the case of reindexing an unmodified (or slightly modified) document from the original source.

Allowing iteration of the modified data is awkward - either we need specialised iterator subclasses to handle this, or we flatten the changes onto the old data to give the current representation when asked to iterate.

Marking for 1.2.x as this probably doesn't require ABI or API changes.

[Olly Betts]

Not a blocker for 1.4.0.

[Olly Betts]

The "adding a single term" case will need to rewrite the termlist table entry, which I think means this is inevitably O(number of terms indexing that document) (at least until we support replace_document() on termlist-less databases).

It may still be significantly faster despite that unchanged O() though.