Relative cost error

[stephane-caron]

A proposal from @bodono in bodono/scs-python#63 (comment) is that tolerances should be set relatively (e.g, an objective error should be relative to the true objective).

For the Maros-Meszaros test set all true objectives are $>0$ (except perhaps HS51?), so we can use the relative variation instead:

$$ rel\ cost\ error := \frac{|f(x_{solver}) - f(x^*)|}{|f(x^*)|} $$

This isn't entirely satisfactory though, because on some problems in different test sets $f(x^*)$ could be zero. 🤔