bottom-up corrected: prevent leakage while keeping cross-connections

[ilkerkesen]

This is even better than the previous. Previously, cross-connections used to cause multi-modal information leakage from the contracting branch to the expanding branch, and to prevent this, I had developed two different versions named as bottom-up and bottom-up (disconnected) for the ablations study. In this version, there's an additional visual contracting branch which does not contain any linguistic information, and expanding branch uses outputs produced by this branch in cross-connections excluding the last one.