`realize_reduction` requires a typed kernel for comparative operations

```python
import loopy as lp

knl = lp.make_kernel(
"{[i]: 0 <= i < n }",
"""
z = max([i], a[i])
""",
lang_version=(2018, 2)
)
knl = lp.realize_reduction(knl)
```

dies [here](https://github.com/inducer/loopy/blob/main/loopy/library/reduction.py#L202=) because the `dtype` is `None`.

I'm guessing this wants a symbolic smallest/largest number node that is made concrete after type inference. Are there any similar examples of things that remain abstract until after type inference/later on in codegen?