[MLIR] CSE misses an optimization opportunity to eliminate redundant `memref.load`

MLIR `--cse` misses an optimization opportunity to eliminate a redundant `memref.load` within the same region, even when the first load (at line 3) strictly dominates the second (at line 6) and no intervening writes or side effects exist.

### Input Program

**input.mlir**

```llvm
module {
func.func @domination(%arg0: memref, %arg1: index) -> (i32, i32) {
%0 = memref.load %arg0[%arg1] : memref
cf.br ^bb1
^bb1:
%1 = memref.load %arg0[%arg1] : memref
return %0, %1 : i32, i32
}
}
```

### Command
```bash
mlir-opt --cse input.mlir
```

Version: 157af9cb0e9cad7df0e8e1d44c78f08ffa7c48e7

Start by running `mlir-opt --cse input.mlir` on the reproducer in the issue and inspect the resulting IR. Trace the MLIR CSE handling for the two `memref.load` operations, using the dominance and side-effect conditions described here. Done means the redundant second load is eliminated and the returned values remain correct.

Written by the indexing model from the issue text.