[llvm] [FunctionAttrs][IR] Fix memory attr inference for volatile mem intrinsics (PR #122926)

[Eli Friedman via llvm-commits]

https://github.com/efriedma-quic commented:

The primary value of fixing the attributes on the declarations would be to reduce the likelihood of mistakes: if someone accidentally queries the attributes on llvm.memcpy directly, instead of going through the CallBase wrappers, they get the conservatively right result.  Otherwise it would be basically the same as this patch.

I'm not sure I see the value in splitting all the memory intrinsics into volatile/non-volatile versions; that seems like a lot of redundancy.  I guess most places don't use the names of memory intrinsics directly anyway, so maybe it's not that bad.

https://github.com/llvm/llvm-project/pull/122926