Use incremental decompression

[seanlinsley]

Currently when decompress is called, all data in that row is fully decompressed and recombined into the original data structure, and then returned as a Vec. If the user immediately consumes the data and reduces it into a smaller result, those unnecessary allocations can be a significant portion of the runtime and in an extreme case with a very wide struct could cause an OOM.

After #15, we can re-implement the internal decompression code to run in batches of 256 rows using ChunkDecompressor

[seanlinsley]

There may be a marginal performance improvement by retaining a Vec<QueryStat> across each incremental decompression and passing that to the caller with clone() on the individual row, thereby avoiding the large Vec allocation each iteration. But that's just a theory; the simpler version that constructs a new Vec each iteration and mutably pop() each row to pass to the caller might be just as fast.