Remove worker reconnection logic

[gjoseph92]

It's proving difficult to merge a fix for #5480. #6329, #6272, and #6341 are causing tests to fail in unexpected and hard-to-debug ways. Because reconnection is quite broken as things stand right now (and causing issues for users #6228), instead of fixing it, how about we just remove it?

If a worker's connection to the scheduler is broken, have it shut down immediately. This would at least puts us in a consistent state, instead of deadlocking.

Questions:

1. If there's a nanny, should the worker restart?

cc @fjetter @crusaderky @mrocklin

[fjetter]

I'm in favour of a clean shutdown. I suggest to log an informative message, e.g. suggesting users to increase distributed.comm.timeouts.tcp if this issue occurs frequently.