Different precisions for data across platforms

[ageorgou]

Some datasets are written using either 32 or 64 bits in different environments. From memory, this seemed to depend on the user's operating system rather than the Python version. 32 bits are enough for the datasets in question, so this needlessly increases the size of the NWB files. Changing it would also simplify the signature checking, as we currently have to cast specific datasets and attributes to the "right" type to maintain consistency in signatures.

We need to check whether the difference occurs when reading the data (should be easily fixable) or when pynwb writes it. If it's the latter, we should see how/whether we can control it.

See also #7.