Proposal: Do not attempt recovery of network timeouts

[jackc]

Currently, pgx uses net.SetTimeout to implement context cancellation and it attempts to recover from the network timeouts when possible.

There are three different distinct cases that are or were handled:

1. A timeout occurs during a Write resulting in a partial write. We always consider this a fatal error and close the connection in this case as there is no way to reliably recover.
2. A timeout occurs before or during a Write resulting in 0 bytes written. Previously, we attempted to recover from this case. However, this case still breaks a TLS connection (and worse it is not reported properly: crypto/tls: permanently broken tls.Conn should not return temporary net.Error golang/go#29971) So we now always consider this a fatal error as well.
3. A timeout occurs during a Read. Because of the underlying message buffering it should always be safe to recover from these errors. pgx does attempt to recover from these.

Trying to recover from timeouts is a very complex and error prone part of the code. (#494, #506, ....)

I propose abandoning the effort and considering all errors on Read or Write fatal. This should have no effect on anyone using the connection pool aside from removing a source of heisenbugs. The only impacted use case I can see is using a single connection with context cancellations and needing to recover and continuing to use that connection after a cancellation. That seems to be very unusual case -- and it never could be done entirely reliably since some timeouts always were fatal.

[fcantournet]

yes please :)