Performance Bug: Scan rel is slow when bound node offsets are random

[ray6080]

Kùzu version

master

What operating system are you using?

No response

What happened?

Currently, we assume scans of rel tables are sequential (bound node offsets are sequential), and cache the whole node group header (CSR lengths and offsets) ahead of actual scans. However, in cases like recursive joins and graph algorithms, the access pattern can be random and we lost the guarantee of sequential scans.

Under these cases, we should take in a batch of bound node offsets (with the assumption that they are all sorted? this is to be decided by the caller, but this should be a reasonable gurantee to be provided), and have a smarter way of cacheing only necessary offsets and lengths from the node group header.

The optimization should be all done inside initializeScanState without the need of introducing another set of interfaces.

Are there known steps to reproduce?

No response