Fix bug with handling of null values in dictionaries

[adriangb]

Currently given the query json_get_text(col, 'a') on the data ['{'x': 0}', '{'x': 0}', '{'a': 1}'] where col is a dictionary encoded column originally with keys [0, 0, 1] and values ['{'x': 0}', '{'x': 1}'] we return a dictionary with keys [0, 0, 1] and values [null, null, 1].
But if you look at how arrow-rs builds up dictionaries they always put the nulls in the keys, not the values. The spec does not require this, but I think that things in arrow-rs or DataFusion assume it is so (based on panics I've seen in prod).
This PR works around those bugs elsewhere while we investigate them further.

[adriangb]

Unfortunately:

1. This seems hard to reproduce in pure DataFusion.
2. This implementation is buggy and it's hard to implement a generic dictionary rebuild function; ideally we could just use the <Type>DictionaryBuilders.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.02%. Comparing base (999d672) to head (643294d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #70      +/-   ##
==========================================
+ Coverage   82.80%   83.02%   +0.22%     
==========================================
  Files          15       15              
  Lines        1128     1143      +15     
  Branches     1128     1143      +15     
==========================================
+ Hits          934      949      +15     
  Misses        132      132              
  Partials       62       62              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[davidhewitt]

Let's add a test case to show that we return nulls in the dictionary keys in cases when we would produce null values.

I wonder if we should switch to using dictionary builders immediately if we're going through the effort to re-pack dictionaries at the end anyway.

[adriangb]

I pushed a test and simplified to just set keys to null as you suggested. I can confirm the test used to fail.