Short-living Memory Context in the optimiser

Hi,

Here is a rebased and slightly commented Tom's patch separated from the 
discussion [1].

I still occasionally see that planner peak memory consumption is 
triggered by selectivity estimation of massive arrays and clauses 
generated for multiple partitions.
I think that in case of a growing number of selectivity estimations, the 
planner should consume memory in some scalable manner, so this thread is 
an attempt to draw community attention to this issue.

I question some decisions in that patch. For example, field 'depth' and 
even 'usage' can be incorporated inside the MemoryContext structure. 
This can make the reset function safer. To avoid overhead, it may be 
used only for the 'short-living' class of memory contexts, which can be 
declared with an additional parameter of AllocSetContextCreate on a 
context creation.

[1] https://www.postgresql.org/message-id/1367418.1708816059@sss.pgh.pa.us

-- 
regards, Andrei Lepikhov