[llvm] [AArch64] Override isLSRCostLess, take number of instructions into account (PR #84189)

[Graham Hunter via llvm-commits]

huntergr-arm wrote:

rebased to address merge conflicts.

I thought I'd seen a regression in bwaves from spec, but it turns out that it's performance is just unstable, at least on the machine I was benchmarking on. So I think we see almost no performance difference from this patch, but it does remove a few instructions from some loops. I'll commit as is, but we'll probably want to revisit it the ordering later without the strict focus on vscale-relative offsets that I had when originally writing this.

https://github.com/llvm/llvm-project/pull/84189