Document the semantics of floating-point atomic fetch_min / fetch_max

For floats, whether the operation is:

```cpp
std::max(old, new);
```

or

```cpp
std::max(new, old)
```

matters. We don't say, but we should say something about this at least.

Since we are doing a `old.max(new)`, I think we should document this as `std::max(old, new)` for consistency.