Neither of these terms is appropriate for a bug in an arcade game that was programmed in assembly language and runs without benefit of memory-protection hardware or operating system.
"Undefined behaviour" is a term-of-art in C and related languages, coined by the C standards committee back in 1989. Code has undefined behavior when the language specification doesn't define what it will do. There is no such thing in Z80 assembly language: the effect of every opcode with every possible input is well-defined. The conventional English meaning of "undefined behavior" can be read to apply -- the kill screen is behavior not defined by the people who wrote the game -- but I wouldn't use it in this context because it's too likely to give the wrong impression.
"Segmentation fault" is a term-of-art in POSIX, derived ultimately from PDP system programming jargon. Segmentation faults happen when a program attempts to access a memory address that isn't "mapped" to anything: the hardware and operating system detect this and shut down the malfunctioning program, in a carefully-defined way that allows the program a chance to recover. Something like this could have happened as a result of a bug in the Pac-Man game program, because the Pac-Man circuit board only populates a little less than half of the Z80's 64kB address space with ROM, RAM, and peripherals, but I haven't been able to find out what the real hardware would do if the software attempted to access unmapped memory. Whatever it would do, though, it would be inappropriate to describe as a "segmentation fault", because the "operating system" for Pac-Man (to the extent it even has one) is not an implementation of Unix and, again, it would give the wrong impression.
The level 256 bug, meanwhile, does not access unmapped memory, so it's moot.
It is accurate to say that the game has a bug which manifests upon advancing to level 256. It is also accurate to say that the root cause of the bug is an integer overflow, and that its consequences are memory corruption (or, equivalently, violations of memory and type safety). These are all general-purpose CS terms defined without reference to any particular language or OS environment.
It's also accurate to observe that the effects of the bug are similar to the effects, within a modern environment, of memory-corruption bugs that don't provoke segmentation faults. If you read any of the Project Zero exploit writeups, you will see a remarkable similarity to Don Hodges' analysis of the Pac-Man kill screen.
Note that an emulator that doesn't faithfully reproduce the kill screen when fed the Pac-Man ROMs is not emulating the game hardware correctly.