It's much easier and cleaner to subdivide quads to refine shapes when modeling. For example, you can split the quads along an entire edge to get a new clean edge for manipulation (ex. to bevel it). If you try to do the same with triangles, you get a jagged mess.

I believe also quads also deform better during animation.

This might be a naive/stupid question, but wouldn't it be relatively easy to merge triangles in the same plane into polygons automatically? (I suppose few triangles from this process would be in the same plane maybe?)

If they’re actually in the same plane, yes. But that’s rarely the case except in hard surface modeling.

Besides, merging them unintelligently still doesn’t necessarily form clean edge loops, which are required for good model topology.

They are not. Polygons are a terrible representation since unlike triangles they do not cleanly represent a unique planar surface. With more than 3 points you will always have an ambiguity (or several) about which (numerical) plane corresponds to the actual face. For some graphics applications this may or may not matter much, but it is very important for anything using the mesh for physical computation.

Co-planar quads will always subdivide into two coplanar tris. That's the crux on why the modeling work flow works. The GPU is going to turn it into triangles anyway as long as a few fundamental rules with indices are up kept, so you're mostly getting the best of both worlds here.