[SROA] Large array causes SROA pass OOM

- **Problem:** `AggLoadStoreRewriter` runs before SROA partitioning and recursively splits first-class aggregate loads and stores into scalar accesses. For a very large array, this can create a pathological amount of intermediate IR before later optimization passes can simplify it.

- **Example:** Given a 256 MB aggregate copy:

```llvm
%value = load [268435455 x i8], ptr %src, align 1
store [268435455 x i8] %value, ptr %dst, align 1
```

Running SROA first can attempt to create `extractvalue`, `getelementptr`, and `store` operations for every array element, and it will be OOM.

I try to reorder the pipelines(#210875), which run `MemCpyOpt` before SROA, but i think it's not the best fix, so open an issue.

Start by tracing AggLoadStoreRewriter in the SROA pass and reproduce the 256 MB aggregate load/store example. Compare the intermediate IR and memory behavior when MemCpyOpt runs before SROA; done means avoiding pathological per-element IR creation without introducing incorrect transformations.

Written by the indexing model from the issue text.