Currently, the usage of the streamer library seems to make things less stable memory-wise than before - which goes against the goal of the library. We should achieve higher number in tpch_concurrency roachtest with the streamer enabled. I think the root cause is that the streamer creates many more objects in general (increasing pressure on the GC), and large slices are not accounted for yet.

We should:

• remove allocations as much as possible
• make sure all things are accounted for
• consider making Result an interface so that depending on Hints and Mode we could have smaller-size implementations.

Jira issue: CRDB-16320