Encoding version 2: binary-only memory layout

[willdealtry]

This PR adds an additional data encoding that is entirely binary in terms of the essential data structure descriptors. The aim is to make encoding and decoding faster, have a storage structure that can be entirely described by a set of POD structures (located in memory_layout.hpp), and pave the way for a more sophisticated approach to data encoding that will follow in a separate (smaller) PR.

The main work in this PR is to remove protobuf structures entirely from the internal implementation of data encoding and compression, and provide a mapping layer which can translate from standard C++ structures to either the legacy (protobuf) format or the new binary one.

The binary encoding is a much more direct representation of the in-memory structures, which fall into three main groups. The structures around EncodedFieldCollection describe the layout of a (usually compressed) field in storage, although a fixed set of optional EncodedFields are used at the head of the segment to bootstrap the decoding. StreamDescriptor describes features such as the names, types and dimensionality of the data being represented (thus it is oriented primarily towards the features of data in memory, as opposed to data in storage. Finally the TimeseriesDescriptor represents elements specific to segments that describe and reference sets of other segments, such as indexes etc, i.e. it contains data relevant to time series and dataframes as a whole rather than to their component parts.

[IvoDD]

I've paied closest attention to the protobuf mappings and new memory layout to check it all looks compatible. It all makes sense and most of my comments are just questions to double check everything makes sense.