[llvm] [VPlan] Fix `pseudoprobe` being treated a vector intrinsic (PR #185347)

[David Sherwood via llvm-commits]

david-arm wrote:

Yes, it is indeed strange. `getVectorIntrinsicCost` calls `getVectorIntrinsicIDForCall` to get the ID, then asks for the cost of the vector form of llvm.lifetime.start, etc. This cost is then used by `setVectorizedCallDecision` to decide on the correct strategy. However, we don't actually vectorise `llvm.lifetime.end` and it remains in a single scalar form. I think that means that `getVectorIntrinsicCost` is calculating the wrong cost, since it doesn't reflect the actual output.

https://github.com/llvm/llvm-project/pull/185347