State of copying in the core

[shachlanAmazon]

This ticket is intended to follow the number of copies of the data sent from the wrapper before being sent on the wrapper. Some of the steps here can potentially be optimized away in order to improve performance.

• copy from socket to an owned buffer
• copy from rotating buffer into protobuf owned Chars. This might not be a copy - requires investigation.
• copy from protobuf buffers into Cmd
• Copy in cmd on calls to get_packed_command. This can be removed by using a single-copy structure, as proposed here.
• In cluster mode, .command() is called on routables, and it allocates vec and copies the first part of the command.
• In cluster mode, on multi-shard commands, new commands are created from the original command.
• In the parser, the full command is copied into a buffer before being passed to the tcp stream. This is probably unnecessary, since the parser doesn't do any parsing on inputs, only on outputs.