IRCE: regression b/w LLVM 18 and 19 due to "Apply loop guards" in loop constrainer

https://godbolt.org/z/hd46Eq1xb
Test: `opt -passes=irce` on
```
define i32 @test(i32 %x) {
bb:
%add = add nuw i32 %x, 16
%icmp = icmp slt i32 %add, %x
br i1 %icmp, label %bb2, label %bb1

bb1: ; preds = %bb4, %bb
ret i32 0

bb2: ; preds = %bb4, %bb
%phi = phi i32 [ %add5, %bb4 ], [ %add, %bb ]
%icmp3 = icmp ult i32 %phi, 1
br i1 %icmp3, label %bb4, label %bb7

bb4: ; preds = %bb2
%add5 = add i32 %phi, 1
%icmp6 = icmp slt i32 %add5, %x
br i1 %icmp6, label %bb2, label %bb1

bb7: ; preds = %bb2
ret i32 0
}
```
IRCE is not able to handle range check from this case starting version 19. The reason of this lies here:
https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/Utils/LoopConstrainer.cpp#L91

```
auto StartLG = SE.applyLoopGuards(Start, L);
auto BoundLG = SE.applyLoopGuards(BoundSCEV, L);
```

This makes the predicates unprovable in some cases. Remove this code (just take `StartLG = Start` and `BoundLG = BOundSCEV`)and it will become provable.

There are also cases when these loop guards to help, off course. But this change is at least controversial.

The workaround I can propose is to factor out this code and try to prove with and without loop guards. Alternatively, investigate why SCEV cannot prove it.