cache k-ANN graph in memory

Let's keep the graph in memory. Like, in TABLE_SHARE.
Add nodes there as they're read from the table, e.g. on searches.
Update on inserts.

When it'll grow beyond some (configurable) limit, start removing nodes that are close to each other. Or remove the bottom level in HNSW. The goal is to still keep enough in memory to find any row with no more than one table access.

Not clear how to do transaction isolation.