I'm not sure what to recommend as an optimal learning ISA, though. Ideally something that's mostly RISC, with a small but unified memory model, and no virtual memory, and which has all the math ops that embedded RISC ISAs never bother to implement. (In the end, probably no real architecture ever made that exact set of compromises, and so it'd have to be something made up — some kind of bytecode abstract machine ISA designed specifically to be used in education, an ISA equivalent to languages like Logo or Scratch, or OSes like Minix.)
In practice, I think I would recommend learning assembler by programming for the original Gameboy's LR35902 ISA (which is almost, but not exactly, the Z80/8080 ISA). Not because it's the optimal learning ISA per se, but because it gives you a complete machine including graphical video output, with a compact unified address space (no separate IO-register read/write ops), modern tools that target it, and good emulators that almost always include built-in debuggers and memory viewers — where, as well, the memory is small enough to get it all on-screen and have room left over.
Oh, and unlike other ISAs, where people either never hand-write code for them, or where no hand-written code survives in a readable form, with LR35902 we have several complete projects with hand-written assembly codebases (or, well, reverse-engineered + annotated disassemblies of originally hand-written assembly — just as good!) around to be studied. And, of course, they're just the sorts of things kids interested in computers would be interested in studying the workings of: video games! (https://github.com/pret/pokered is a shining example here.)