Fix BGReader deadlock when stopping background reads

[sergeydobrodey]

BGReader.Read could block forever if Stop() was called while the background goroutine was stuck in a read.

Read now falls back to direct reading when status is StatusStopping, not just StatusStopped.
Adds a unit test that reproduces the deadlock and verifies the fix.

Tests
go test ./pgconn/internal/bgreader -run TestBGReaderReadDoesNotWaitWhenStopping

Possible it's a fix for #2119

[jackc]

I think a race condition would be possible with this change. BGReader is defined as concurrency safe, but there is no guarantee the underlying io.Reader is. With this change the underlying io.Reader can be used concurrently.

1. BGReader.Start() is called. bgRead() is started in a goroutine and blocks on Read() from the underlying io.Reader.
2. BGReader.Read() is called and execution waits at r.cond.Wait()
3. BGReader.Stop() is called. This sets status to StatusStopping, but it does not stop the goroutine running bgRead().
4. BGReader.Read() is called again. Because status is StatusStopping it also calls Read() from the underlying io.Reader.

In practice, this shouldn't happen, both because we don't call BGReader.Read() concurrently, and we only use it with net.Conn which is concurrency safe. However, I don't want to introduce a latent bug either.

[sergeydobrodey]

Agree. I'll close this PR