[MLIR][CSE] Missed `memref.load` elimination across dominated block

MLIR CSE currently misses a simple redundant `memref.load` optimization opportunity across a dominated block. In the following example, the second `memref.load` at line 10 loads from the same memref and the same index as the first load at line 4. The block containing the second load is strictly dominated by the first load, and there are no intervening operations that may write to the memref or otherwise invalidate the loaded value.

### Input Program

**input.mlir**

```llvm
module {
func.func private @composable()
func.func @dominated_block(%arg0: memref, %arg1: index, %arg2: i1, %arg3: i1) -> (i32, i32) {
%0 = memref.load %arg0[%arg1] : memref
cf.cond_br %arg2, ^bb2, ^bb1
^bb1: // pred: ^bb0
call @composable() : () -> ()
cf.br ^bb2
^bb2: // 2 preds: ^bb0, ^bb1
%1 = memref.load %arg0[%arg1] : memref
return %0, %1 : i32, i32
}
}
```

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

Version: ef88536bbf62469eb2396d719d237b3191cf4843