fix writer async close

[achille-roussel]

This PR simplifies and fixes internal synchronization in the kafka.Writer type.

There were a few issues:

• marking the writer as closed was done via an atomic operations in Close, racing with the check in WriteMessages
• increments of the partitionWriter wait group was done asynchronously after starting goroutines, racing with calls to Wait on these wait groups

I removed the sync group in partitionWriter and instead managed the count of inflight operations globally on a single wait group in Writer; we only care about the writer's Close call to block until all inflight operations are done, there is no need to have a hierarchical synchronization scheme.

Fixes #778

[egaldamez]

Hello, this is an interesting fix. I'm currently facing an issues in which I'm trying to Close a writer so that buffered messages would be flushed and sent to my cluster. However, this doesn't always happen. My scenario is as follows:

1. I have a go application (v1.7) that accepts http requests and these requests are transformed into kafka messages (buffered by the application and flushed every 15 seconds or 1000 messages)
2. I use vegeta to send traffic to this application @ 1000 msg/s
3. I stop vegeta. At this point, my application has some messages buffered.
4. I send an OS interrupt signal to my application (which is configured to listen to these signals)
5. I expect that the messages buffered are flushed, but this does not always happen especially when I am sending a high volume of messages.

So I'm wondering if this could be because of the race conditions addressed in this PR. I'm looking forward to fetching the latest version and testing this for myself :) thank you.