[arrufat]

As discussed in #3176, this PR adds tfidf_transformer, which learns TF-IDF weights from sparse vectors of term counts and turns documents into TF-IDF vectors that the sparse linear trainers can use directly.

I needed it for a page classifier that trains dlib's linear SVM on hashed n-grams, where it replaced our hand-written TF-IDF code with identical results. The defaults follow scikit-learn's TfidfTransformer (smoothed idf, raw counts, unit length).

In a separate commit, I also added normalize() and normalize_inplace() for sparse vectors, since the transformer needs it. Let me know if you'd rather not have it, and I'll drop it.

The only doubt I have is around the name normalize_inplace: I followed make_sparse_vector_inplace, but maybe there's a better one.

Closes #3176

[arrufat]

If you're ok with it, I might submit a couple more PRs after this one:

• feature hashing for text n-grams, using the existing murmur_hash3
• a text classification example showing how all these pieces work together

[davisking]

Nice, this is a great PR.