[llvm] [AMDGPU][GlobalISel] Improve odd size s16 vector store legalization (PR #206342)

[Jiaqi He via llvm-commits]

heturing wrote:

The update currently only fixes the case for `s16`. However, I observed similar behaviour for `s8` as well (e.g., `<5 x s8>`). Do we also want to add the case for `s8`?

https://github.com/llvm/llvm-project/pull/206342