Queued publishes are failed with connection shutdown errors when connection keep-alive fails

[Bret Ambrose]

I'm not sure if this is intentional or not, but I find it confusing. Loosely, it looks all the publish tasks are getting run as cancelled when the keep alive failure tears everything down. Is there a better error? Is there a more coherent sensible offline publish strategy that we should pursue?

This was discovered based on aws/aws-iot-device-sdk-python-v2#62

[Dengke Tang]

#154