Let's assume a setup where two of these I2C-bridges are connected to the same I2C bus and on the end of each bridge there is an identical I2C slave. It's obvious that you need some way to distinguish between the two slaves - as they are identical hardware they have the same address and would answer both to a request to one of them.
In "pass through" mode, you set up an address translation table that remaps the addresses of the slaves. In this way you set up one bridge to change the (virtual) address of the slave from the actual X to Y, and set up the other bridge to change it from X to Z. This is nice if you have to communicate with both slaves in an interleaved way, but requires you to first set up this table (which you only can do if you know the addresses in advance) and your master needs to keep track about the changed addresses.
In "pass through all" mode, no addresses are changed. The advantage is, you can communicate with any slave on the other side of the bridge as if it was in your local system and you don't need to set up a translation table. Obviously, you can only set one of the two bridges to this mode at a time, otherwise you would get some garbage response if they answer at the same time.
This is handy if you want to communicate with several slaves on one bridge at the same time and don't need any communication with slaves on the other bridge.