Empty unlimited chunked variables cause crash

[bnlawrence]

If ones has created a variable which is intended to be chunked, but it is currently empty, when the file is read, we get a stack dump that ends with this:

    def _read_node_header(self, offset, node_level):
        """ Return a single node header in the b-tree located at a give offset. """
>       self.fh.seek(offset)
E       ValueError: cannot fit 'int' into an offset-sized integer

This pull request includes code to create a file which manifests the problem, a test to expose it, and a fix.

[bmaranville]

I think this will work - it is also possible to test if the chunk address is UNDEFINED_ADDRESS, which will happen when no data has been written to the Dataset yet (see change in usnistgov/jsfive@f228420 , which I should have backported to pyfive)

EDIT - I think maybe the test for UNDEFINED_ADDRESS is important here because sometimes you encounter datasets with non-zero shapes but which have not been written yet (initializing a dataset and writing data to it are two separate steps).

[bnlawrence]

That seems sensible. At this point I'm not minded to follow through fixing this here, as where it gets done in the new H5D.py will be slightly different. What I have there is:

# look out for an empty dataset, which will have no btree
if np.prod(self.shape) == 0 or dataobject._chunk_address == UNDEFINED_ADDRESS:
    self._index = {}
    return

(This is in the context of caching the b-tree when we instantiate a DatasetID, which we do when we create a variable instance with eg. `x=myfile['variable']. We do that at this point so that all threads in a thread pool have their b-tree before they get going on their bit of work.)