Make termsim matrix positive definite even with negative similarities

[Witiko]

When a SparseTermSimilarityMatrix object is constructed with non-default parameters that satisfy the following constraints: symmetric and positive_definite and threshold < 0, the matrix embedded in the constructed object may not be positive definite. This PR hardens the code against negative term similarities, making sure that the term similarity matrix is always positive definite.

[Witiko]

Note that this set of parameters is unusual, since it makes sense to threshold the similarities at zero (that is why we are using a sparse matrix afterall), so although this bugfix changes the produced matrices, few users should be affected.

[mpenkov]

Thank you for this PR. Unfortunately, I'm not very familiar with the underlying code, so it's difficult for me to comment on the substance of your PR. I do notice that it appears to be a bugfix. So, I have a few questions:

• Is it possible to add a unit test? Is it worth it?
• How does the bug manifest itself? In other words, what can happen if we don't fix this?

[Witiko]

Is it possible to add a unit test? Is it worth it?

It is possible to unit-test this specific bug. This will be useful if someone reverts the fix in the future, or reimplements the SparseTermSimilarityMatrix class.

How does the bug manifest itself? In other words, what can happen if we don't fix this?

If this bug is not fixed, the term similarity matrix produced by the SparseTermSimilarityMatrix class constructor will not always be positive definite, violating the API contract.

Gensim does not depend on the positive definiteness of a term similarity matrix at the moment. However, a user may produce word embeddings from the matrix using Cholesky factorization. This will fail for matrices that are not symmetric positive definite.

As I noted above, this bug manifests itself rarely. The users affected by the bug will be mainly those who use grid search or a similar technique to optimize the threshold parameter.

[mpenkov]

Ok, could you please add a test to this PR?

[Witiko]

I have just added the unit test in c87aa31.

[mpenkov]

OK, this looks good to me. Thank you for your contribution.