Missed optimization: fold gep i8 (gep double %ptr, %x), %x * 24 into gep double %ptr, %x * 4

```llvm
define i1 @src(i64 %arg0, ptr %arg1) {
%v0 = getelementptr double, ptr %arg1, i64 %arg0
%v1 = mul nuw nsw i64 %arg0, 24
%v2 = getelementptr i8, ptr %v0, i64 %v1
%v3 = load double, ptr %v2, align 8
%v4 = call noundef double @llvm.fabs.f64(double %v3)
%v5 = fcmp ogt double %v4, 0x10000000000000
ret i1 %v5
}

define i1 @tgt(i64 %arg0, ptr %arg1) {
%scaled_idx = mul nuw nsw i64 %arg0, 4
%v0_opt = getelementptr double, ptr %arg1, i64 %scaled_idx
%v3 = load double, ptr %v0_opt, align 8
%v4 = call noundef double @llvm.fabs.f64(double %v3)
%v5 = fcmp ogt double %v4, 0x10000000000000
ret i1 %v5
}

```

Available in alive2 when time limit = 50000:

```llvm
define i1 @src(i64 %arg0, ptr %arg1) {
#0:
%v0 = gep ptr %arg1, 8 x i64 %arg0
%v1 = mul nsw nuw i64 %arg0, 24
%v2 = gep ptr %v0, 1 x i64 %v1
%v3 = load double, ptr %v2, align 8
%v4 = fabs double %v3
assume_welldefined double %v4
%v5 = fcmp ogt double %v4, 0.000000
ret i1 %v5
}
=>
define i1 @tgt(i64 %arg0, ptr %arg1) {
#0:
%scaled_idx = mul nsw nuw i64 %arg0, 4
%v0_opt = gep ptr %arg1, 8 x i64 %scaled_idx
%v3 = load double, ptr %v0_opt, align 8
%v4 = fabs double %v3
assume_welldefined double %v4
%v5 = fcmp ogt double %v4, 0.000000
ret i1 %v5
}
Transformation seems to be correct!
```
godbolt: https://godbolt.org/z/Wjr9qanvd
Pattern found in: https://github.com/dtcxzyw/llvm-opt-benchmark/blob/main/bench/g2o/optimized/line3d.ll