[InstCombine] Check stride preconditions in `foldCmpLoadFromIndexedGlobal` (#234187)
When folding a comparison of a load from a constant global into a
comparison on the index in `foldCmpLoadFromIndexedGlobal`, the
enumeration of offsets `ConstOffset + Idx * Stride` within the global is
assumed to be exhaustive. This holds as long as the stride is positive
for nusw GEPs, and it is a power of two for non-nusw GEPs, as `Idx *
Stride` may wrap. Neither assumption is guaranteed by the decomposed
linear expression, so make sure these are checked explicitly.
Fixes: https://github.com/llvm/llvm-project/issues/233092.