New form for variational tk (as in PRB paper) seems to work much better than the previous one (that was from the thermalk code), it is also twice as fast as the same matrix can be reused, but some more testing is needed. In particular, I have some doubt about the index to use in freqm1