sql: cache ImmutableTableDescriptor

[yuzefovich]

Extracting the second observation from #42770:

2. We always construct a new ImmutableTableDescriptor. These things are meant to be cached. I'm not sure how best to plumb a cache (or which one).

cc @ajwerner as I think you've been thinking about this and possibly there already exists a better issue.

Jira issue: CRDB-3608

[ajwerner]

I started to hack on this in #53747. I'll work to get that over the finish line. I think that that should enable us to cache even more inside the ImmutableTableDescriptor.

[postamar]

We now do effectively cache these in the descriptors collection. The exception to the rule are the table descriptors based on the protos in the distsql specs, which ultimately could be replaced by an (ID,Version) pair used to look them up in the descriptors collection.