`show_doc` and `_non_empty_keys` fail when input default value is an array

Hi, I opened an [issue in the fastcore repo](https://github.com/fastai/fastcore/issues/489) but after digging into the problem, it may be better suited here.

As commented there, when you a function with a default numpy array input like:

```
def foo(c: np.array = np.arange(4)
):
'''
Test function with numpy default
'''
```

you end up with the following error.

```
---------------------------------------------------------------------------
[ ..... other unnecessary error info ......]

File ~\miniconda3\lib\site-packages\nbdev\showdoc.py:25, in (.0)
---> 25 def _non_empty_keys(d:dict): return L([k for k,v in d.items() if v != inspect._empty])

ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

```

The problem arises in `_non_empty_keys`, when this function checks whether the default value of the given variable is empty with `v != inspect._empty`. When the default `v` is a `np.array` of size bigger than one, the previous operation outputs a `np.array`, which makes that the if statements fails, as shown here:


When working only with arrays, using `.any()` or `.all()` solves the issue. However, in this case is not applicable as `v` can be of any type. See that for most of the other defaults, the inequality outputs a `bool`. I don't exactly know what would be the preferred way to solve this issue, that works for any `v` type.

The traceback points to showdoc.py and _non_empty_keys, used by show_doc; start there and reproduce the issue with the NumPy-default function shown. Check how empty defaults and array defaults are distinguished when comparisons may not return a scalar. Done means show_doc handles the example without ValueError while preserving existing behavior for other default types.

Written by the indexing model from the issue text.