Improve safety of NER build

[cthoyt]

This PR updates the NER index build to skip entries whose norm terms are none or contain an empty string. This shouldn't happen in practice, but depending on the pipeline used to construct terms, this might happen. A few alternatives to consider:

1. Have the Term class raise errors when None is given for norm_text
2. Raise an error in the NER index build

[bgyori]

This is actually tricky to solve perfectly. We already throw a ValueError if a Term's text evaluates to False in the constructor (https://github.com/gyorilab/gilda/blob/master/gilda/term.py#L53-L54). So I think checking norm_text the same way could make sense. Still, depending on how the terms are loaded, this could be inconvenient. We could also check for norm_text.strip() I suppose since a norm_text consisting of just spaces doesn't make sense either.