Provide peekInt8 to reduce allocations

[shanson7]

Background:
In a project I use, sarama produces by far the most ongoing allocations (mainly, I think, because have lots of small messages) which negatively impacts our GC stats. I recently went through and tried to reduce them where I could and have been quite successful in that sarama doesn't show up in the pprof top10 by alloc_objects. This is the simplest change I made so far. The others follow the pattern from #1161 and are more invasive.

I benchmarked this change by consuming 5 million messages and then dumping out a heap profile.
Before the change:

Showing nodes accounting for 30777165, 99.24% of 31012468 total
Dropped 6 nodes (cum <= 155062)
Showing top 10 nodes out of 19
      flat  flat%   sum%        cum   cum%
  10362760 33.41% 33.41%   25777082 83.12%  github.com/Shopify/sarama.(*MessageBlock).decode
   5172719 16.68% 50.09%   25887933 83.48%  github.com/Shopify/sarama.(*MessageSet).decode
   5123816 16.52% 66.62%    5123816 16.52%  github.com/Shopify/sarama.(*partitionConsumer).parseMessages
   5087542 16.40% 83.02%    5087542 16.40%  github.com/Shopify/sarama.(*realDecoder).peek

Here you can see that peek is responsible for 16.4% of the objects allocated, a little over one per message (makes sense, because of the record batching in kafka).

After the change:

Showing nodes accounting for 25577632, 99.73% of 25647802 total

Peek doesn't show up at all and we allocate 25.6 million objects vs 31 million, which is the expected reduction.

[varun06]

If this change work, do we still need peek method?

[shanson7]

There are no existing uses of peek. Should I remove it?

[varun06]

let's wait. @bai you have any inputs?

[shanson7]

Any update?

[bai]

Thanks for your contribution!