Improve `reader` error messages

0.4: reduce memory footprint (segmentio#547)

fix zstd decoder leak (segmentio#543)

* fix zstd decoder leak

* fix tests

* fix panic

* fix tests (2)

* fix tests (3)

* fix tests (4)

* move ConnWaitGroup to testing package

* fix zstd codec

* Update compress/zstd/zstd.go

Co-authored-by: Nicholas Sun <olassun2@gmail.com>

* PR feedback

Co-authored-by: Nicholas Sun <olassun2@gmail.com>

Remove deprecated function (NewWriter) usages (segmentio#528)

0.4 broker resolver (segmentio#526)

* 0.4: kafka.BrokerResolver

* add kafka.Transport.Context

* inline network and address fields in conn type

Fix for panic when RequiredAcks is set to RequireNone (segmentio#504)

* Fix panic in async wait() method when RequiredAcks is None

When RequiredAcks is None, the producer does not wait for a
response from the broker, therefore the response is nil.
The async wait() method was not handling this case, leading
to a panic.

* Add regression test for RequiredAcks == RequireNone

This new test is required because all the other Writer tests use
NewWriter() to create Writers, which sets RequiredAcks to
RequireAll when 0 (None) was specified.

Fix consumer group timeout handling (segmentio#510)

PR segmentio#509 introduced a bug that could create infinite rebalance loops
when the Timeout is less than the RebalanceTimeout because the JoinGroup
call could timeout before the rebalance completes.

consumer groups: ensure deadline is set on broker interactions (segme…

…ntio#509)

Since the Conn does not have a timeout by default, it's possible for a
connection to hang if the remote side crashes before responding.  This
PR ensures that every interaction with the broker has a deadline so that
the consumer group won't get stuck if the coordinator goes down.

Merge remote-tracking branch 'origin/master' into 0.4

ReadBatchWith bug fix and new MaxWaitTime option (segmentio#482)

This fixes a bug where ReadBatchWith could incorrectly return `io.EOF`
in certain `errShortRead` scenarios.  Since `io.EOF` means the batch
was handled correctly, this could cause consumers to continue to use
a connection that is in a bad state.

This also adds a MaxWait option that can be used to improve the
way that the conn deadline and the fetch request interact.  Instead
of trying to calculate a correct timeout and change the max wait
time, this allows the caller to explicitly configure both the wait
and the deadline independently of one another.