limit memory consumption of scan interface

[skyzh]

Currently we use the same code path as vectored get, which limits on the number of keys we get, but we removed that limit for the scan interface.

There could be multiple possible implementations:

• Gather statistics for the metadata keyspace and split it into several scans.
  • Might be easy to implement, still reuse the vectored get code path
  • Statistics could be inaccurate
• Implement k-merge based iterators
  • Need a new implementation for scans
  • Guaranteed memory consumption: delta buffer size per file * num of files involved in the scan