You would be surprised: these guys tend be extremely frugal with their topology and only put vertices where they absolutely need to be, relying on the subd algorithms to do the heavy lifting. Small models load faster, are easier/faster to fix & modify: optimization becomes a virtuous cycle. As mentioned above: it is all about productivity. IIRC, Merida’s full-body skin control cage is something like a measly 25k quads, compared to the 125k triangles of a Battlefield 3 soldier.
With that said, cartoon animation is its own niche ; ‘photoreal’ VFX characters will often have much higher surface complexity to support the needs of hyper-detailed articulation (muscle/wrinkle/vascular deformations).
Definitely not as a general practice. The vast majority of assets are 100% Catmark subds (including all the cars in Cars) and rendered as such (you can find papers on how prman intersects those for ray-tracing). You may find some Loop subds for simulation assets like cloth. And some straight low-res poly models for sets background objects (ex. the trash mountains in Wall-E).
Hard/complex-surface modelling with polys is positively medieval. For general animation & gaming, Catmark (with crease support) is generally your best choice. There is a fairly good choice of apps (i include Mudbox & ZBrush here, since they both use subds under the hood), but pretty much all of them come with serious issues & require extensive customization for advanced users. For CAD, NURBS are currently your only hope (maybe one day we will have NURCCs…)