Optimise Threshold Tracer queues/sorting

The heap sorting process uses up a lot of time in TLT, and seems to be one of the largest factors in where time is spent when tracing. This is more pronounced when using Python client, where lots of extra spans are produced and potentially accumulate.

See https://issues.couchbase.com/browse/PYCBC-507?focusedCommentId=287932&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-287932