Make -q :: faster by calculating interactions on the fly instead of pre calculating them

[olgavrou]

This PR affects only -q :: for now, could be extended to also work for --cubic :: in different PR.

We don't hold all possible interactions anymore but calculate them on the fly by keeping namespaces when they are found.
This should not affect explicit quadratics when passed

Could potentially improve more. Reduces run time for a ccb adf run with a file with 347k examples and 3 interactions (Action,Slot,User) from ~ 11m41s to ~ 1m50s

[jackgerrits]

Looks great! Just a thought, should we test combining -q :: with explicit -q AB and other interactions --interaction AAA --interaction :::?

[olgavrou]

Looks great! Just a thought, should we test combining -q :: with explicit -q AB and other interactions --interaction AAA --interaction :::?

Good idea about a mixed quadratics test, will add

[lalo]

Relevant #1527

[lalo]

We will need to update these docs (see bottom) https://github.com/VowpalWabbit/vowpal_wabbit/wiki/Feature-interactions

[olgavrou]

We will need to update these docs (see bottom) https://github.com/VowpalWabbit/vowpal_wabbit/wiki/Feature-interactions

Awesome thanks, I would not have noticed that

Added more tests to this, leaving it broken for now (tests timeout needs to be increased) since I am experiencing something strange for the latest tests: -q :: -interactions AAA --interactions :::. The interactions generated match exactly with those when run from master (seen from --invert_hash output) but their weights are ever so slightly different and I don't know why and if it should be a concern or not

[rajan-chari]

Looks good.