The Rx data replaces the Tx data in USISRL. Thus if USICNT=8, after 8 SCKs (->USIIFG), the Tx data has been overwritten with the Rx data.
For the nRF24L01P, the first Tx byte (command) "trades" with the status register from the chip. The second and following (as needed) Tx bytes trade with the command response (if any).
Since the register read/write commands are (mostly) 2-byte, you could probably play tricks there with USI16B. Such tricks wouldn't of course scale easily to the multi-byte registers or e.g. read/write-fifo commands.