[llvm] [SLP][AMDGPU] Fix load-address typo in v4f16 combine tests (PR #222112)

[Pablo Reble via llvm-commits]

================
@@ -465,30 +465,23 @@ bb:
 ; FIXME: Should always vectorize
----------------
reble wrote:

Not yet.

`gfx90a` fully vectorizes both `copysign_combine_v4f16` and `canonicalize_combine_v4f16` but `gfx8/gfx908/gfx1030` still only combine them pairwise. The FIXME is still accurate for those three targets, so I'd leave it in until that's addressed.

https://github.com/llvm/llvm-project/pull/222112