Number of gradient evaluations not applicable to scipy's gradient-free optimizers

Since commit [#28a474e](https://github.com/google/jaxopt/commit/28a474e307665971ce3e535d1984325e0ec5b1e4), scipy optimizers that do not use gradients do not work from `jaxopt.ScipyMinimize` anymore.

Because `res.njev` is accessed on the result regardless of the `method` argument, you get an

```
AttributeError: njev. Did you mean: 'nfev'?
```

if you run `jaxopt.ScipyMinimize` e.g. with Nelder-Mead or Powell.