You always want to compile to the easiest target language that solves the problem well and allows you to do the optimizations you want to do. By compiling to C++, they can focus on static typing and closures etc., but don't have to worry about generating machine code, and register allocation and whatnot.
Back in school, I worked on a project to compile Matlab. We picked Fortran as the target, because it was very similar in terms of operating on matrices. But it's typed, and had a few very fast compilers available.
Python compiles to its own bytecode, but only when a script is first interpreted.
Haxe compiles to many different source code languages ahead of time, including C++, Python, Java, JavaScript, and C#. It also compiles to a custom "Neko VM" bytecode.
So, "compiling" isn't necessarily ahead of time, depending on what level of "compilation" you're talking about.
Well, not that implicit, seeing that we also say "JIT compiler".