GitLab

Previously, we'd only retry when a write failed (in case the peer was using
single-use streams). However, in the new world of half-open streams, writes
still succeed on remote-closed streams so this isn't sufficient. Instead, we
need to retry on read failure as well.

Note: this was technically broken before because a peer could write on a stream
before receiving the close message causing the write to succeed but the
subsequent read to fail.