Very high jemalloc memory usage with single doc array index

This bug is related to MB-34674 where GSI buffer usage was optimized to avoid bloat of memory due to appends in collatejson package. After GSI fix, we observed that indexer heap usage came down but plasma jemalloc usage still showed 19G of usage. Repro steps are below:

Repro:
To repro: Create a document that is 1.5MB large. Create an array index with array size over 3000. Resulting size of all entries should be in the tune of 3.5GB for single document. large_covering_field itself (below) should be around 1.5MB.
CREATE INDEX idx ON default(ALL array, large_covering_field)