Because you can? It looks like a great project for getting started with nontrivial FPGA design.
Programming-wise I'd say full FPGA is less useful than QMK. Doing a direct 1:1 mapping from key inputs to USB HID report isn't too bad to do in an FPGA, but dynamic behavior like macros, layers, leader keys, mod tap, auto shift, and so on are significantly easier to implement in a regular programming language. If you want flexibility you're basically forced to have your FPGA run a soft core, so at that point why not go for a regular MCU?
In theory you could make an argument for lower latency, but that doesn't really apply when you're limited by USB 2's 1000Hz polling rate while some off-the-shelf MCU-based keyboards use USB 3 for 8000Hz polling.
With regards to latency: reaction time is the wrong thing to compare against, since input latency can be noticeable even when it is much smaller than a human's reaction time. And wanting to have your keyboard latency be no worse than 1ms can make sense given that it's only one of many components of the total input latency of a computer. Reducing mouse or keyboard latency from 8ms (125Hz) to 1ms (1000Hz) is pretty much the low-hanging fruit; it's not as cheap or easy to squeeze 7ms out of the other parts of the chain.
Programming-wise I'd say full FPGA is less useful than QMK. Doing a direct 1:1 mapping from key inputs to USB HID report isn't too bad to do in an FPGA, but dynamic behavior like macros, layers, leader keys, mod tap, auto shift, and so on are significantly easier to implement in a regular programming language. If you want flexibility you're basically forced to have your FPGA run a soft core, so at that point why not go for a regular MCU?
In theory you could make an argument for lower latency, but that doesn't really apply when you're limited by USB 2's 1000Hz polling rate while some off-the-shelf MCU-based keyboards use USB 3 for 8000Hz polling.