ENH: Implement __round__ special method for ndarrays

This would allow calling round(array). While this would certainly be equivalent to np.round(array), a similar feature is already there with abs(array) (which calls __abs__) and np.abs(array).

[seberg]

Sounds like we should just start with decimals=None raising an informative error, that way users can use round and get an informative error when that doesn’t work.

It could still be allowed in some future.

[mhvk]

Partially, this seems to be a question also of whether the current round(numpy_scalar) is desired behaviour. I also do not like scalars and arrays treated differently.

An alternative would be to generalize and have round(array, ndigits=None) cast to int64 (raising or warning on overflow). That would be backwards compatible except in cases like 1e300 where the current answer is already not great.