Use numpy string dtype to get accurate memory usage

[jeromekelleher]

We're currently using object dtypes in the BufferedArray class when dealing with string input and this gives us inaccurate numbers when we look at the array nbytes. To allow us understand where memory is being used (and perhaps use it better) we should switch to the new numpy StringDtype which should help.

So, at least we're buffering the data in an efficient format, and ideally this is basically what Zarr will use ultimately as well.