[llvm] [AArch64][TTI] Reduce cost for splatting whole first vector segment (SVE) (PR #145701)

[Graham Hunter via llvm-commits]

huntergr-arm wrote:

Something seems wrong with the initial costs in the precommit patch; the 512b test (`dup_seg_512b_d`) has a higher cost with native 512b support than for 256b only.

I left out SVE as a feature check here, since NEON only has 128b vectors to start with and only needs N-1 movs to do this, where N is the number of 128b segments.  (Or may not need to do it at all, since the same register could be reused).

https://github.com/llvm/llvm-project/pull/145701