[llvm] [VPlan] Handle WidenGEP in narrowToSingleScalars (PR #166740)

[Ramkumar Ramachandra via llvm-commits]

artagnon wrote:

> > if all operands are single scalar, we should be able to handle this in narrowToSingleScalar, and remove the special handling in the recipe altogether?
> 
> This is currently not possible, as narrowToSingleScalars' users check is too weak.

To prove this, I've rebased onto https://github.com/llvm/llvm-project/pull/166559.

https://github.com/llvm/llvm-project/pull/166740