[llvm] [LoopVectorize] Don't discount instructions scalarized due to tail folding (PR #109289)

[John Brawn via llvm-commits]

john-brawn-arm wrote:

The general problem with vectorization cost calculations is, as I've mentioned before, that everything uses the reciprocal throughput as the cost kind, but we should be using code size when optimizing for code size. I was going to fix that after this, as it's a much larger change, and we'll still need this patch as the cost calculation that it fixes is currently wrong no matter what cost kind is being used, but I'll get to work on getting it done and then get back to this.

https://github.com/llvm/llvm-project/pull/109289