Here's a clue: -
It's all about matching impedances to avoid data corruptions due to incorrectly terminated cables. Take a look at the table circled in red.
With a turns ratio of \$\sqrt2\$ impedances are transformed by this squared. So an impedance of 24.3 ohms for Rt becomes transformed to 48.6 ohms. As the CS8900A outputs differentially the impedance seen by the line (RJ45 side) is double this at 97.2 ohms. Given that there may be some series resistance in the driver outputs of about an ohm the 24.3 ohm might be more like 25 ohms and then double this because it is a diff driver then double because the turns ratio is \$\sqrt2\$ and you get 100 ohms.
To match 150 ohms, divide by 4 to get 37.5 and this is pretty near to the 37.4 ohms specified but what about the 1 ohm I used above. Well there are losses in the transformer that are in-effect parallel resistors (core losses due to eddy currents) and these will be more prevalent at higher impedances so I guess, if I new the exact characteristics of the transformer and could be bothered to look at the tech spec of the line driver for the chip I might be able to justify things a tad better. Gut feeling tells me I'm about right!
So what happens on 3V3 supplies? Well the output drive level is going to be smaller for sure so using a transformer to bump up the output voltage makes sense but you've still got to pay head to the impedances. A 1 to 2.5 steps up impedances by \$2.5^2 = 6.25\$ so the 8 ohm specified becomes 16 ohm differential then 100 ohm on the RJ45 side.