Add profiling support for jobs and framework overhead

[bwhite]

As hadoop is often used to speed up some task, it should be second nature to determine if it is doing a good job or not. Since this spans several languages (Cython, C, Python, Java, and various client code) standard profiling techniques don't work properly. A solution is to have an internal profiling aggregator that can be enabled to produce semantically relevant timing information.

Relevant information

• TypedBytes serialization overhead
• Pyinstaller overhead
• Client Map/Combine/Reduce functions

It is important that this doesn't clutter the codebase and should have minimal performance impact.