This commit moves the responsibility for limiting writes to a socket from nano::socket on to nano::channel_tcp. The intended behavior of nano::socket is to be a thin wrapper around a raw tcp socket with the addition of queueing on a strand plus statistics gathering.
The drop policy is part of the channel abstraction which varies depending on the channel used. channel_udp uses the operating system to drop writes it can't service and channel_tcp needs to track the number of outstanding writes to determine what should be dropped.
The drop policy was being passed all the way through to nano::socket, punching a hole through the nano::socket abstraction.

This commit also removes the concurrency enum as redundant. All socket operations were being serialized through the strand implicitly and the only other function of having a non-concurrent channel was to bypass the socket rate limiter. Non-concurrent users of channel_tcp should manage their own rate limiting if they don't want their buffers dropped. The only users of non-concurrent channels appear to be bootstrapping and this limitation already appears to be respected.

This also removes the need for queueing the buffer and callback in a separate container as they're already captured in objects to be executed on the socket's strand. This removes the need to flush the buffers on socket close.