sql/catalog: rework validation and hydration to use cached descriptors

[ajwerner]

Is your feature request related to a problem? Please describe.
Today validation, including cross descriptor validation, occurs when reading a descriptor from the store in the catalogkv package. This is a low-level layer. The cross-descriptor validation not only validates the references from the descriptor being retrieved to what it references, it also fetches the dependencies one layer deeper so that that all references of descriptors which should back-reference this descriptor are validated.

This can cause real efficiency problems when, say, resolving schemas. Each schema will resolve its database which will, in turn, resolve all of the schemas. This is an O(N^2) proposition.

Describe the solution you'd like

To me, a key problem is the layering between retrieval of "raw" descriptors from the database and its relationship to validation. Instead of fetching and validating in one call, we should fetch and return a builder and then validate using already fetched descriptors in the descriptor collection. This can avoid the need to fetch any additional descriptors from the store most of the time.

Describe alternatives you've considered
We could do less cross-descriptor validation.

Additional context
Relates to #64388 and #64089.

Epic: CRDB-11534

Jira issue: CRDB-11135

[ajwerner]

@postamar if I'm not mistaken, you've now fixed this 🥳

[postamar]

I got part of the way there only. There are still direct descriptor reads which avoid the collection for validation-related lookups, see readValidationDereferencer in catkv. That's the thing we need to get rid of, ultimately.

[postamar]

I have fixed this completely with #77630