Favor replacing backpressure detection heuristic with write buffer threshold callbacks

[vitaly-krugl]

Pika's current backpressure detection mechanism is meant as a heuristic to detect when the broker is blocking the connection. RabbitMQ has for some time supported Connection.Blocked/Unblocked notifications (also supported by Pika via add_on_connection_blocked_callback() and add_on_connection_unblocked_callback()), so at least with RabbitMQ, the latter is preferred.

Backpressure detection in Pika (Connection._detect_backpressure()) is presently slated for deprecation since (at least) RabbitMQ now makes backpressure explicit through Connection.Blocked and Connection.Unblocked notifications.

What may be more useful to users of Pika's non-blocking connections for managing memory usage is a simpler mechanism that notifies them when the connection's write buffer size exceeds a certain user-configured high-water mark (so they may suspend publishing to avoid overwhelming the app's memory) and when it goes below a user-configured low-water mark (so they may resume publishing). This concept is akin to asyncio's WriteTransport.set_write_buffer_limits().

[vitaly-krugl]

cc @lukebakken, @michaelklishin