Split Marshal and Buffer logic

[santazhang]

Marshal: only in charge of (de)serialization, make it .h header only code

• generic interface
• serialization code will be like:

template <class Buffer>
Marshal<Buffer> operator << (Marshal<Buffer>& m, Data& data);

Buffer: handles data storage

• bounded buffer (UDP): ring buffer
• Unbounded buffer: list of chunks as in original Marshal implementation

Note that current python code might need major update to support this design.