[llvm] [LV][AArch64]: Utilise SVE ld4/st4 instructions via auto-vectorisation (PR #89018)

[Eli Friedman via llvm-commits]

efriedma-quic wrote:

Doing deinterleaving as trees sort of makes sense for high interleaving factors... I've seen loops that benefit from deinterleaving with interleave factors as high as 12.  I'm a little concerned the abstraction layers here are going to make cost modeling less accurate, though; ideally, the vectorizer should be able to estimate the cost of an ld4.

https://github.com/llvm/llvm-project/pull/89018