unsafe is useful for a lot more than C interop. For example, doing image analysis with actual pixel data can be horribly slow while in "safe" mode. Flip over to unsafe mode, and doing the same work with direct access to the image buffer, is an order of magnitude faster.
Doing wholesale development in C/C++ for inclusion in a .Net project can certainly make sense, but it all comes down to performance. The classic advice would be to write the entire application in your language of choice, then profile. When you find a critical section of code that takes 50%+ of processing time and cannot be realistically tuned anymore while using a managed language, then re-implement that one routine in C and integrate.
A real life example: I was prototyping a path finding algorithm. After showing it worked with C#, I needed to make it fast enough to run many times a second. I had already built up some nice support UI in C# that showed all of the factors used for routing decisions, the results, etc. So I reimplemented the core part of the routing algorithm in C++. If I had started in C++, I doubt I would have gotten as far.