changed Fetch.default value from 32KB to 1MB

[bshafiee]

In my experiments of consuming very busy topics, I noticed this parameter makes a huge difference. Bumping up the default value from 32KB to 1MB almost increased consumption rate by 5X. 1MB seems to be the default value used in librdkafka as well.

[eapache]

To capture a bit more of our offline conversation here:
I originally chose 32KB because I figured that would be the point of peak performance - any response over that size will be fragmented at the network layer anyway and I assumed that would limit the performance gains from further batching.

Clearly this was false - the cost we reduce through larger batches is not just per-packet routing costs but more per-request latency costs (and also the number of disk reads on the broker).