Leela is not going to fit in an 8-bit machine, period. The NN weights take many megabytes to represent.
It's conceivable that Stockfish's evaluation function could fit - I think it's only comparable in complexity to some of the better 8-bit engines (eg. those by Ed Schröder). However, Stockfish's main strength is that it can do a lot of searching very quickly, which requires large transposition tables - which will not fit.
Most 8-bit chess engines relied on a relatively sophisticated evaluation function and a much less sophisticated search function, which was not able to search very deep in the game tree. That fitted well with the slow CPUs and limited RAM of the time.
Subsequent developments allowed searching deeper and more efficiently using faster CPUs and more RAM; the 68020 and ARM were popular in dedicated chess engines after the 6502 was superseded.
Only very deep searches allow using a simpler, more naive evaluation function, which itself is a tradeoff to further increase search depth. And only the extra RAM of modern machines permits doing the deep, smart searches, even with unlimited time. These are inherently unavailable to vintage hardware.