slowlog: improve accurracy of replay

[undingen]

Running several smaller replays showed that with the following changes the accuracy can be improved:

• use the InnoDB transaction ID as sort criteria (if available)
• sort queries also in non accurate mode

I will run tomorrow further tests and hope to improve it further and will verify that I don't see regressions on large captures.