Well, it does accept fractional polling rates, but you still have to play the song, then precisely calculate what that fractional polling rate should be.
Example: The title theme from Super Mario World.
You need to precisely measure the length of a section of the song. For instance, a loop of the whole song is 38.677458 seconds.
A measure of the song is 3.215750s. So the loop is 12 bars long.
Then you arbitrarily decide how many IT rows you want in a frame. Let's pick 256.
Do some division. 256/3.2231215 = 79.4261091.
That would be your desired update rate.
But then you find that if you add .003 to the update rate, it might align better to the IT file grid.
Anyway, the whole IT dumping thing is a mess and is done the wrong way. It really should just everything with the timestamp at which it happens, then process that information to determine recurring ADSR volume envelopes, frequency sweeps, and other characteristics of the instruments, then determine the real rate they are applied on the grid. Right now it's just using the "Okay, .01 seconds passed, have the channels changed?" algorithm. It has issues with notes being misaligned, or volume or frequency changes immediately before the new note event.