Allow ignoring kwargs inside fn_kwargs during dataset.map's fingerprinting

### Feature request

`dataset.map` accepts a `fn_kwargs` that is passed to `fn`. Currently, the whole `fn_kwargs` is used by `fingerprint_transform` to calculate the new fingerprint.

I'd like to be able to inform `fingerprint_transform` which `fn_kwargs` shoud/shouldn't be taken into account during hashing.

Of course, users should be aware to properly use this new feature, just like the internal usages of `fingerprint_transform` [does](https://github.com/huggingface/datasets/blob/2699593b33ee63d17aad2a2bfddedd38a8df57b8/src/datasets/arrow_dataset.py#L2700).

### Motivation

This is originally motivated by https://github.com/huggingface/transformers/pull/18351#issuecomment-1263588680.

Nonetheless, consider a more general processing function that accepts a kwarg that does not influence it's output:
```python
def fn(example, verbose=False):
...
```

Then `dataset.map(fn, verbose=True)` would not benefit from dataset caching.

I'm not sure if other methods in the `Dataset` API could benefit from this feature.

### Your contribution

Based on `fingerprint_transform `'s `wrapper` function [here](https://github.com/huggingface/datasets/blob/c59cc34fcd2a369d27b77cc678017f5976a926a9/src/datasets/fingerprint.py#L443), it seems to me that it should be possible to make `.map`/`._map_single` accept something like `fn_use_fingerprint_kwargs`/`fn_ignore_fingerprint_kwargs` (probably another arg name). This would then be used by `fingerprint_transform.wrapper` to better/more flexibly hash the transformation.

I could contribute with a PR if this feature and approach look good to you.