It depends on the final effect you want to achieve. In my opinion it's much better to work on single models, and then import them to ArcGIS with replace model option for single building footprints. This way you will have a multipatch layer with every building as a separate feature. Then you can select single buildings or use the layer for analytical purposes (provided that the multipatch models are closed geometry) eg. selecting all the buildings that are intersecting with some other multipatch (eg. radius of some kind of phenomenon). You can also have per-building attributes stored in the layer. If you have one feature for all the multipatch you can't, as far as I know, divide it to single buildings (AFAIK multipart to single part does not work) and you can have attributes defined only for all the layer (the only feature in the layer).
But if you want to have all the buildings in one model, you should be able to union the extruded footprints into one feature and then export them as one object into collada.
And as a sidenote the only 3D formats that support georeferencing are to my best knowledge: VRML, OpenFlight and collada (version 1.5 - ArcGIS uses 1.4 specification), so the best way to maintain your georeferencing data is by means of GIS layers (building footprints, point shapefiles, multipatches). You can replace certain objects in those layers with your externally edited models (by means of symbolising features with 3D symbols, or by replacing features in multipatch layers with your models).