[MLIR][Linalg] `--convert-arith-to-llvm` rewrites a `linalg.matmul` body and leaves the op unverifiable

## Reproducer

### small.mlir
```mlir
func.func @f(%a: memref<2x3xi32>, %b: memref<3x2xi32>, %c: memref<2x2xi32>) {
linalg.matmul ins(%a, %b : memref<2x3xi32>, memref<3x2xi32>) outs(%c : memref<2x2xi32>)
return
}
```

### To reproduce:

```
mlir-opt small.mlir -o /dev/null # parses and verifies
```
```
mlir-opt small.mlir --convert-arith-to-llvm
```

### Crash:

```
small.mlir:2:3: error: expected add/mul op in the body
linalg.matmul ins(%a, %b : memref<2x3xi32>, memref<3x2xi32>) outs(%c : memref<2x2xi32>)
^
small.mlir:2:3: note: see current operation:
"linalg.matmul"(%arg0, %arg1, %arg2) <{indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], operandSegmentSizes = array}> ({
^bb0(%arg3: i32, %arg4: i32, %arg5: i32):
%0 = "llvm.mul"(%arg3, %arg4) <{overflowFlags = 0 : i32}> : (i32, i32) -> i32
%1 = "llvm.add"(%arg5, %0) <{overflowFlags = 0 : i32}> : (i32, i32) -> i32
"linalg.yield"(%1) : (i32) -> ()
}) : (memref<2x3xi32>, memref<3x2xi32>, memref<2x2xi32>) -> ()
```

Start with the provided small.mlir reproducer and run mlir-opt both with and without --convert-arith-to-llvm to compare verification behavior. Trace the conversion's handling of the linalg.matmul body; done means the pass no longer leaves an unverifiable linalg.matmul operation and the reproducer completes successfully.

Written by the indexing model from the issue text.