Optimize memory usage of bulkloader

[xiangzhao632]

This PR optimize memory usage of reduce stage of bulkloader while not harming speed。
Related to #4529

---

This change is 

Docs Preview:

[pawanrawal]

Hi @xiangzhao632

Thanks for the PR. Could you please merge dgraph:master into your fork so that the diff is cleaner. Also, could you share some benchmarking numbers that you were able to observe before and after this change? Things like the memory bulk loader used and the time that it takes to load your dataset.

[xiangzhao632]

@pawanrawal Thanks very much for reply! I have merged dgraph:master into my fork. And I have finished my test. The number is :
About Data: 300million rdf (600million edge). The total memory is 192G. Before change: load time 38m01s, memory usage 67G; after: load time 14m40s, mem 18G.
I noticed that since v20.03.1, the reduce stage of bulkloader has been changed from a min-heap style to a partitionkey based method, done by f7d0371. My optimization is based on the min-heap version, so I switch back.
We have loaded about 50 billion rdf(about 83 billion edges) in 15 hours using the bulkloader on one 192G and ten 128G machines.

[xiangzhao632]

memory usage before change(total memory is 192G)

and the load time

[xiangzhao632]

memory usage after change

and the load time

[awsl-dbq]

cool

[xiangzhao632]

Hi @pawanrawal, has anyone tested this? Could someone tell me this pr works or not?
OOM is a big issue, for two reasons:

1. The largest memory of my machines is 192g (many other teams are 128G). Such memory can only cope with 3 billion rdf, while my team have more than 50 billion rdf. I know some teams who want to try dgraph, but due to the OOM issue, they can't init a cluster by bulkloader and they turned to liveloader.
2. Less memory usage in the reduce stage means that --reducer can be set to a larger value, which can also increase the loading speed.
   Another issue Inconsistent bulk loader failures #5361 about this.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[mangalaman93]

@xiangzhao632 Thank you for the PR. We are happy to look at the PR again if you can sign the CLA and raise it against the main branch. If you need any help, please let me know.

[github-actions]

This PR has been stale for 60 days and will be closed automatically in 7 days. Comment to keep it open.