Rule for atomic_pointerset

[willtebbutt]

TuringLang/DynamicPPL.jl#799 re-highlighted a problem that causes Mooncake.jl to fail in specific situations in DynamicPPL (specific VarInfos). Happily, this just appears to be a problem with the rrule!! for atomic_pointerset, not anything more widespread. See https://github.com/TuringLang/DynamicPPL.jl/actions/runs/13218146414/job/36899908042?pr=799 . In particular, it looks like I basically hadn't considered the possibility that you might be setting a pointer-to-a-pointer when I wrote the rule, and therefore failed to consider that zero_tangent would not work. Moreover, this is implementation is simply incorrect for pointers (they have fdata), so it is a good thing that it errors loudly.

The fix would be to modify atomic_pointerset to account for the fact that the thing being pointed to might have fdata associated to it, and to not construct a zero_tangent if so. I should probably also check this case for pointerset, as I imagine that I made the same mistake there.