BOW not showing attributes in sparse

[ajdapretnar]

Describe the bug
A bit tricky bug to describe. There are two underlying issues:

• Bag of Words can return all 0 features.
• Data Table in sparse is not showing all nan features (probably it can't or maybe word=nan).

To Reproduce
Say we have the following documents:

	cat	dog	sleep
Cat sleeps.	1	0	1
Dog sleeps	0	1	1
Cat sleeps, dog sleeps.	1	2	1

When computing TF-IDF for "sleep", the IDF is 0. There are three ways of computing IDF.

1. math.log10(number_of_docs / number_of_docs_with_word) (how we do it)
2. math.log10(1 + number_of_docs / number_of_docs_with_word) (how we do it with Smooth IDF)
3. math.log10(number_of_docs / (number_of_docs_with_word + 1)) (how it is recommended)

How scikit does it: idf(t) = log [ n / df(t) ] + 1 or idf(t) = log [ (1 + n) / (1 + df(t)) ] + 1 if smooth = True.

To reproduce: Create Corpus (with above docs). Bow (TF-IDF). Data Table.

Expected behavior

1. Resolve the nan result in bow. Why do we not use scikit-learn? Should we reconsider how we compute IDF?
2. Resolve the display of nan in sparse.

Orange version:
3.37.0

Text add-on version:
1.7.0

[ajdapretnar]

Related to #1069.