Improve performances on depth > 0

Improving performances by adding cache on Serializer.to_representation (see investigation results here: https://docs.google.com/document/d/1DOBKF825Pb4H77Fl5JesJq9FoNtYMRnfe_OB-U974TM/edit?ts=5f311be3#heading=h.wvkrq25rofkc)

Gain is from 25% to 75% faster (see perf_result.csv on source code)

Will fix applications/hubl#748 (closed)