>This is the only place where I saw this kind of requirement.

It's been literally the definition since the 1980s.

If the source and target map are not assumed to be surjective submersions then you need to build into your definition the requirement that for a Lie groupoid $X_1 \rightrightarrows X_0$ the pullback $X_1\times_{s,X_0,t} X_1$ is a manifold. Charles Ehresmann for instance took this point of view in the 1950s. But then, many other constructions break without assuming source/target are submersions as you need to pull these back all the time. Submersions are also the class of maps of manifolds that are the 'saturation' of the singleton Grothendieck pretopology consisting of coprojections $\coprod_\alpha U_\alpha \to M$ of open covers, and the whole point of Lie groupoids is that they present differentiable stacks on the site of manifolds and open covers. This is not something that was considered before the late 1980s/1990s. The paper _[Groupoids in categories with pretopology](http://www.tac.mta.ca/tac/volumes/30/55/30-55abs.html)_ by Meyer and Zhu in TAC gives the general theory of groupoids internal to categories with a class of maps like that of submersions. It is a fluke of the category of sets that (split) surjections have all the required properties and so no further constraints are needed. A similar thing happens in the world of Deligne–Mumford stacks, where one assumes groupoids in schemes have étale source and target, or Artin stacks, where the source and target are required to be smooth maps of schemes.

>Are there natural examples of Lie groupoids where source/target maps are not surjective submersions?

Not really, because they are useless. You can come up with some kind of topological groupoid where the object and arrows spaces admit manifold structures, and the source, target etc maps are smooth, but... why?