[llvm] [AMDGPU][SILoadStoreOptimizer] Try to find common base for L/Ss with 0 offset (PR #71126)

[Jeffrey Byrnes via llvm-commits]

jrbyrnes wrote:

We should also prefer base with offset 0.

For example, if we have
```
    addr1 = &a
    addr2 = &a + 1024
    addr3 = &a + 2048
    addr4 = &a + 3072
```
`promoteConstantOffsetToImm` will choose `addr4` as anchor. However, we can express all the addresses using an anchor of `addr` and that involves no extra computation for producing the base. I'll address this separately though.

https://github.com/llvm/llvm-project/pull/71126