Refactor intermediate result (de-)serialize methods

[pykello]

This is a preparation step for adding alternative intermediate result formats. The idea is that all serialization/deserialization for intermediate results should go through the same file so we don't need to modify multiple places to add support for new result formats.

IntermediateResultEncoder is the main interface. To use it, first we create an encoder using IntermediateResultEncoderCreate(), then for each row we call IntermediateResultEncoderReceive() . Finally we call IntermediateResultEncoderDone(). The Receive and Done calls may return a string info which needs to be flushed to the output medium (connection or file).

In this PR we don't refactor worker_partition_protocol.c and corresponding CopyTaskFilesFromDirectory() since (1) refactoring them out turned out to be difficult (2) we are going to remove task-tracker in near future anyway.

An example of another result format added using this interface is https://github.com/citusdata/citus/pull/3480/.

[codecov[bot]]

Codecov Report

Merging #3490 into master will decrease coverage by 0.01%. The diff coverage is 95.00%.

@@            Coverage Diff             @@
##           master    #3490      +/-   ##
==========================================
- Coverage   91.56%   91.54%   -0.02%     
==========================================
  Files         179      180       +1     
  Lines       35452    35474      +22     
==========================================
+ Hits        32460    32476      +16     
- Misses       2992     2998       +6     

[marcocitus]

I think an alternative approach could be to have a layering that is more like DestReceiver -> Encoder (binary, text, ..) -> Byte sink (e.g. file, broadcast, ...).

I think the current PR is an improvement regardless, but maybe something worth considering for the future.