Optimize protobuf serialization

[seanlinsley]

Per this flamegraph (from an internal benchmark for the pg_query Rust crate), most of the time is spent serializing and deserializing the protobuf structure.

The straightforward options to improve this are:

• move from message_pack to message_pack_to_buffer, removing the need to call message_get_packed_size
• move to an arena allocator to avoid the repeated allocations and frees unpack to stack protobuf-c/protobuf-c#154 (comment)

Breaking changes might be:

• remove extra nesting in the protobuf definition, like with the abstract Node containing a Option<node::Node> enum
• remove nesting altogether, returning a flat array of nodes that include index references to their children
• have the Rust crate directly manipulate the C data structures instead of using protobuf

And non-breaking (but not straightforward) changes might be:

• move the Rust logic to C so there's no need to access the protobuf in our hot paths