Query classifier does not classify XA transactions correctly.

There is a fast custom parser for detecting transaction boundaries etc. There is also a sqlite3 based parser that handles that and more. The former is used when only the minimal amount of information must be extracted and the latter when it is known that full parsing will be needed.

The latter parser does not correctly classify XA transactions, which will cause transaction boundaries to be missed, which is of relevance if transaction replay is used as it means that the transaction replay buffer will not be flushed when it could be.