Optimize message serialization

Optimize message serialization

There is a MessageBuffer class, which stores a list of byte slices as
std::list<std::vector<unsigned char>> and can be used for serializing
structs. Every member calls a Write method, which creates a new vector
and appends it to the list. After the list is built, a vector with an
exactly right size is allocated and the data is copied. Also, the class
contains unnecessary mutable state, because the logic is shared with
streaming deserialization.

I have replaced the serialization methods with a single function, which
serializes all objects twice. The first pass ignores the data and only
computes the message size, which allows the second pass that actually
writes the data to only use a single allocation. The new interface is
also simpler and more robust.

Change-Id: I6091b71083997faf9302ad8549ade467deb23a58