Do CRT filters actually need to buffer an entire frame, or is that just commonplace because it's the easy way to do things when you're implementing them in software? It seems like if FPGA based upscaling with CRT emulation is in scope, it shouldn't need to delay the signal by more than a few scanlines.

You don't need to add that much latency if the display is capable of some kind of GSYNC-like mechanism and a high data rate.