[llvm] [AArch64] Keep v16i8 -> v2i64 partial_reduce fixed-length for VL > 128 (PR #204938)

[Benjamin Maxwell via llvm-commits]

MacDue wrote:

I think you can mark this PR as "ready to review". The changes LGTM, though I think the description and PR title need updating as they don't describe the actual issue. I think the root issue was splitting the scalable container vector type is not equivalent to splitting the fixed vector it contains. 

E.g., vscale = 2, have a `nxv4i32` container for `v4i32`, splitting the container gives two vectors of `nxv2i32`, which each hold four elements. That's not the same as splitting the underlying fixed vector into two v2i32 vectors. 

https://github.com/llvm/llvm-project/pull/204938