Missed optimization: fold a chain of getelementptr operations with the same index into a single getelementptr with a multiplied index

```llvm
define ptr @src(i64 %arg0, ptr %arg1) {
%v0 = getelementptr inbounds double, ptr %arg1, i64 %arg0
%v1 = getelementptr inbounds double, ptr %v0, i64 %arg0
%v2 = getelementptr inbounds double, ptr %v1, i64 %arg0
%v3 = getelementptr inbounds double, ptr %v2, i64 %arg0
ret ptr %v3
}

define ptr @tgt(i64 %arg0, ptr %arg1) {
%idx = mul i64 4, %arg0
%v3 = getelementptr inbounds double, ptr %arg1, i64 %idx
ret ptr %v3
}
```

godbolt: https://godbolt.org/z/jEqE81K71
alive2: https://alive2.llvm.org/ce/z/CSfwcA

Pattern found in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/openblas/optimized/dgemv_n.ll

Start with the LLVM IR reproducer in the issue and compare its output with the linked Godbolt and Alive2 results. Review the referenced openblas optimized/dgemv_n.ll benchmark pattern to understand the motivating case. Done means the repeated getelementptr chain is recognized as equivalent to the single multiplied-index form without changing semantics.

Written by the indexing model from the issue text.