I will explain the RX side, the TX side is basically the same.
small_hb_dec is the short filter which works at the higher rate. There are 2 of them instantiated, one for I and one for Q. It has 7 taps. One of those taps is the center tap which only requires a shift and not a multiply, and 2 of those taps are zeros. That leaves 4 taps. The taps are symmetric, which leaves 2 multiplies per output. Since we have at least 2 cycles to produce each output, we can use a single multiplier.
hb_dec is 2nd halfband filter and it works at the lower rate. There are 2 of them instantiated, one for I and one for Q. It has 31 taps. One of those taps is the center tap which only requires a shift and not a multiply, and half of the remainder are zeros. That leaves 16 taps. They are symmetric, so that means we need to do 8 multiplies to produce each output. There are at least 4 cycles to produce each output, so we need to do 2 multiplies at a time.
One of those multipliers does the "outer" coefficients, meaning the ones at the very beginning and end of the impulse response, and one does the "inner" coefficients, meaning the ones around the center of the impulse response. This division is purely an implementation choice, and does not affect the output. I could have put the odd ones on one mult and the even ones on the other, or any other split you could imagine. It doesn't matter.