behavior of NDBuffer.as_scalar

[d-v-b]

@brokkoli71 could you explain the logic of NDBuffer.as_scalar?

I noticed that it's doing some data type-specific stuff (there's a special path for np.datetime64). Why was this necessary?

I'm asking because this function is erroring over in #2874 when I add timedelta data types, because timedeltas are not caught in the datetime special case. But I find the need for dtype-specific special cases here suspicious.

For reference, if I replace that function with something much simpler:

    def as_scalar(self) -> ScalarType:
        """Returns the buffer as a scalar value"""
        if self._data.size != 1:
            raise ValueError("Buffer does not contain a single scalar value")
        return self.as_numpy_array().item()

I get errors in a different place, but it looks like those errors are revealing real indexing bugs.

[d-v-b]

update: array.item() returns a native python value, but we want a numpy scalar, which we can get instead via array[()]. I implemented this over in #3027

[brokkoli71]

I don't remember exactly, but there were some typing/mypy problems when just returning self.as_numpy_array().item().
But returning array[()] seems like the way to go anyway, thanks for fixing this.