Re: We need to rethink relation cache entry rebuild

I wrote:
> Basically I think we have to fix this by ensuring that an error escape
> can't occur while a relcache entry is in a partially rebuilt state.

Attached is a draft patch for this. In addition to fixing the stated
problem, it also takes care of a thinko that I found along the way:
RelationClearRelation assumes that any rd_indexprs or rd_indpred trees
can be freed by deleting the rd_indexcxt context, as the comments in
rel.h imply. But actually the code that loads those fields was putting
the trees directly into CacheMemoryContext, meaning that a cache flush
on an index that has expressions or a predicate would result in a
session-lifespan memory leak.

I think this is not too complicated to back-patch --- if anything the
logic is simpler than before. Comments?

regards, tom lane