Context

Re-implement buffer.h in terms of PyByteArray in order to avoid the big memory copy at the end of BSON encoding. The C api methods would return bytearray instead of bytes.

Spike gist: https://gist.github.com/blink1073/3735f9428ce29faca19ab1d901f05529
Spike draft PR: https://github.com/blink1073/mongo-python-driver/pull/12

TLDR; using PyByteArray saves 1× the document size in peak memory usage, while giving a modest performance increase.

Definition of done

Implement the change and verify there are no C correctness issues. Verify on a perf patch build that there are no perf regressions.

Pitfalls

Correctness and performance.