What TFA refers to as latency is the time required to inject a MIDI event/message received by RtMIDI into the perl async object - basically inter-thread wakeup. You certainly expect that to be very, very fast - it's what context switching on any multitasking/multithreaded system relies. "Fractions of a millisecond" is in the right ballpark, though on some systems "Tens of microseconds" would be more appropriate as a description.
What "latency" normally means in the context of MIDI or audio is the delay between the user generating a MIDI event somehow and when the audio corresponding to that MIDI event (e.g. a note) is audible. This is typically measured in relatively small numbers of milliseconds (e.g. 1-30).
These are not really connected in any meaningful sense.