Implement memory bound merging of aggregation results

[nickitat]

We already have a mechanism to reliably bound memory consumption during partial aggregation phase (when we consume table rows and accumulating per-thread aggregation states). This mechanism is called "external aggregation" and the mentioned memory bound is provided by the user through the max_bytes_before_external_group_by setting. Currently maintained aggregation states are flushed to disk once this limit is reached.
During merging we don't have such guarantees: the amount of data residing in memory is proportional to total_aggregation_state_size * (aggregation_memory_efficient_merge_threads / MAX_BUCKETS), MAX_BUCKETS value is constant and equals 256.
This fact complicates running CH in memory constrained environment.

Use case

Aggregation with large state that doesn't fit in memory limit during merging phase.

Describe the solution you'd like

Implement merging algorithm similar to the one we use in aggregation in order.

---

• MBM on initiator when doing distributed aggregation in order Memory bound merging for distributed aggregation in order #40879
• MBM for ordinary aggregation (currently under development)